xref: /freebsd/sys/kern/vfs_bio.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * this file contains a new buffer I/O scheme implementing a coherent
30  * VM object and buffer cache scheme.  Pains have been taken to make
31  * sure that the performance degradation associated with schemes such
32  * as this is not realized.
33  *
34  * Author:  John S. Dyson
35  * Significant help during the development and debugging phases
36  * had been provided by David Greenman, also of the FreeBSD core team.
37  *
38  * see man buf(9) for more info.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/mount.h>
54 #include <sys/mutex.h>
55 #include <sys/kernel.h>
56 #include <sys/kthread.h>
57 #include <sys/proc.h>
58 #include <sys/resourcevar.h>
59 #include <sys/sysctl.h>
60 #include <sys/vmmeter.h>
61 #include <sys/vnode.h>
62 #include <geom/geom.h>
63 #include <vm/vm.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_pageout.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_map.h>
71 #include "opt_directio.h"
72 #include "opt_swap.h"
73 
74 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
75 
76 struct	bio_ops bioops;		/* I/O operation notification */
77 
78 struct	buf_ops buf_ops_bio = {
79 	.bop_name	=	"buf_ops_bio",
80 	.bop_write	=	bufwrite,
81 	.bop_strategy	=	bufstrategy,
82 	.bop_sync	=	bufsync,
83 };
84 
85 /*
86  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
87  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
88  */
89 struct buf *buf;		/* buffer header pool */
90 
91 static struct proc *bufdaemonproc;
92 
93 static int inmem(struct vnode *vp, daddr_t blkno);
94 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
95 		vm_offset_t to);
96 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
97 		vm_offset_t to);
98 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
99 			       int pageno, vm_page_t m);
100 static void vfs_clean_pages(struct buf *bp);
101 static void vfs_setdirty(struct buf *bp);
102 static void vfs_vmio_release(struct buf *bp);
103 static int vfs_bio_clcheck(struct vnode *vp, int size,
104 		daddr_t lblkno, daddr_t blkno);
105 static int flushbufqueues(int flushdeps);
106 static void buf_daemon(void);
107 static void bremfreel(struct buf *bp);
108 
109 int vmiodirenable = TRUE;
110 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
111     "Use the VM system for directory writes");
112 int runningbufspace;
113 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
114     "Amount of presently outstanding async buffer io");
115 static int bufspace;
116 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
117     "KVA memory used for bufs");
118 static int maxbufspace;
119 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
120     "Maximum allowed value of bufspace (including buf_daemon)");
121 static int bufmallocspace;
122 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
123     "Amount of malloced memory for buffers");
124 static int maxbufmallocspace;
125 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
126     "Maximum amount of malloced memory for buffers");
127 static int lobufspace;
128 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
129     "Minimum amount of buffers we want to have");
130 static int hibufspace;
131 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
132     "Maximum allowed value of bufspace (excluding buf_daemon)");
133 static int bufreusecnt;
134 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
135     "Number of times we have reused a buffer");
136 static int buffreekvacnt;
137 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
138     "Number of times we have freed the KVA space from some buffer");
139 static int bufdefragcnt;
140 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
141     "Number of times we have had to repeat buffer allocation to defragment");
142 static int lorunningspace;
143 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
144     "Minimum preferred space used for in-progress I/O");
145 static int hirunningspace;
146 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
147     "Maximum amount of space to use for in-progress I/O");
148 static int dirtybufferflushes;
149 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
150     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
151 static int altbufferflushes;
152 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
153     0, "Number of fsync flushes to limit dirty buffers");
154 static int recursiveflushes;
155 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
156     0, "Number of flushes skipped due to being recursive");
157 static int numdirtybuffers;
158 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
159     "Number of buffers that are dirty (has unwritten changes) at the moment");
160 static int lodirtybuffers;
161 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
162     "How many buffers we want to have free before bufdaemon can sleep");
163 static int hidirtybuffers;
164 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
165     "When the number of dirty buffers is considered severe");
166 static int dirtybufthresh;
167 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
168     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
169 static int numfreebuffers;
170 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
171     "Number of free buffers");
172 static int lofreebuffers;
173 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
174    "XXX Unused");
175 static int hifreebuffers;
176 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
177    "XXX Complicatedly unused");
178 static int getnewbufcalls;
179 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
180    "Number of calls to getnewbuf");
181 static int getnewbufrestarts;
182 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
183     "Number of times getnewbuf has had to restart a buffer aquisition");
184 
185 /*
186  * Wakeup point for bufdaemon, as well as indicator of whether it is already
187  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
188  * is idling.
189  */
190 static int bd_request;
191 
192 /*
193  * This lock synchronizes access to bd_request.
194  */
195 static struct mtx bdlock;
196 
197 /*
198  * bogus page -- for I/O to/from partially complete buffers
199  * this is a temporary solution to the problem, but it is not
200  * really that bad.  it would be better to split the buffer
201  * for input in the case of buffers partially already in memory,
202  * but the code is intricate enough already.
203  */
204 vm_page_t bogus_page;
205 
206 /*
207  * Synchronization (sleep/wakeup) variable for active buffer space requests.
208  * Set when wait starts, cleared prior to wakeup().
209  * Used in runningbufwakeup() and waitrunningbufspace().
210  */
211 static int runningbufreq;
212 
213 /*
214  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
215  * waitrunningbufspace().
216  */
217 static struct mtx rbreqlock;
218 
219 /*
220  * Synchronization (sleep/wakeup) variable for buffer requests.
221  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
222  * by and/or.
223  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
224  * getnewbuf(), and getblk().
225  */
226 static int needsbuffer;
227 
228 /*
229  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
230  */
231 static struct mtx nblock;
232 
233 /*
234  * Lock that protects against bwait()/bdone()/B_DONE races.
235  */
236 
237 static struct mtx bdonelock;
238 
239 /*
240  * Definitions for the buffer free lists.
241  */
242 #define BUFFER_QUEUES	5	/* number of free buffer queues */
243 
244 #define QUEUE_NONE	0	/* on no queue */
245 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
246 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
247 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
248 #define QUEUE_EMPTY	4	/* empty buffer headers */
249 
250 /* Queues for free buffers with various properties */
251 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
252 
253 /* Lock for the bufqueues */
254 static struct mtx bqlock;
255 
256 /*
257  * Single global constant for BUF_WMESG, to avoid getting multiple references.
258  * buf_wmesg is referred from macros.
259  */
260 const char *buf_wmesg = BUF_WMESG;
261 
262 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
263 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
264 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
265 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
266 
267 #ifdef DIRECTIO
268 extern void ffs_rawread_setup(void);
269 #endif /* DIRECTIO */
270 /*
271  *	numdirtywakeup:
272  *
273  *	If someone is blocked due to there being too many dirty buffers,
274  *	and numdirtybuffers is now reasonable, wake them up.
275  */
276 
277 static __inline void
278 numdirtywakeup(int level)
279 {
280 
281 	if (numdirtybuffers <= level) {
282 		mtx_lock(&nblock);
283 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
284 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
285 			wakeup(&needsbuffer);
286 		}
287 		mtx_unlock(&nblock);
288 	}
289 }
290 
291 /*
292  *	bufspacewakeup:
293  *
294  *	Called when buffer space is potentially available for recovery.
295  *	getnewbuf() will block on this flag when it is unable to free
296  *	sufficient buffer space.  Buffer space becomes recoverable when
297  *	bp's get placed back in the queues.
298  */
299 
300 static __inline void
301 bufspacewakeup(void)
302 {
303 
304 	/*
305 	 * If someone is waiting for BUF space, wake them up.  Even
306 	 * though we haven't freed the kva space yet, the waiting
307 	 * process will be able to now.
308 	 */
309 	mtx_lock(&nblock);
310 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
311 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
312 		wakeup(&needsbuffer);
313 	}
314 	mtx_unlock(&nblock);
315 }
316 
317 /*
318  * runningbufwakeup() - in-progress I/O accounting.
319  *
320  */
321 static __inline void
322 runningbufwakeup(struct buf *bp)
323 {
324 
325 	if (bp->b_runningbufspace) {
326 		atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
327 		bp->b_runningbufspace = 0;
328 		mtx_lock(&rbreqlock);
329 		if (runningbufreq && runningbufspace <= lorunningspace) {
330 			runningbufreq = 0;
331 			wakeup(&runningbufreq);
332 		}
333 		mtx_unlock(&rbreqlock);
334 	}
335 }
336 
337 /*
338  *	bufcountwakeup:
339  *
340  *	Called when a buffer has been added to one of the free queues to
341  *	account for the buffer and to wakeup anyone waiting for free buffers.
342  *	This typically occurs when large amounts of metadata are being handled
343  *	by the buffer cache ( else buffer space runs out first, usually ).
344  */
345 
346 static __inline void
347 bufcountwakeup(void)
348 {
349 
350 	atomic_add_int(&numfreebuffers, 1);
351 	mtx_lock(&nblock);
352 	if (needsbuffer) {
353 		needsbuffer &= ~VFS_BIO_NEED_ANY;
354 		if (numfreebuffers >= hifreebuffers)
355 			needsbuffer &= ~VFS_BIO_NEED_FREE;
356 		wakeup(&needsbuffer);
357 	}
358 	mtx_unlock(&nblock);
359 }
360 
361 /*
362  *	waitrunningbufspace()
363  *
364  *	runningbufspace is a measure of the amount of I/O currently
365  *	running.  This routine is used in async-write situations to
366  *	prevent creating huge backups of pending writes to a device.
367  *	Only asynchronous writes are governed by this function.
368  *
369  *	Reads will adjust runningbufspace, but will not block based on it.
370  *	The read load has a side effect of reducing the allowed write load.
371  *
372  *	This does NOT turn an async write into a sync write.  It waits
373  *	for earlier writes to complete and generally returns before the
374  *	caller's write has reached the device.
375  */
376 static __inline void
377 waitrunningbufspace(void)
378 {
379 
380 	mtx_lock(&rbreqlock);
381 	while (runningbufspace > hirunningspace) {
382 		++runningbufreq;
383 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
384 	}
385 	mtx_unlock(&rbreqlock);
386 }
387 
388 
389 /*
390  *	vfs_buf_test_cache:
391  *
392  *	Called when a buffer is extended.  This function clears the B_CACHE
393  *	bit if the newly extended portion of the buffer does not contain
394  *	valid data.
395  */
396 static __inline
397 void
398 vfs_buf_test_cache(struct buf *bp,
399 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
400 		  vm_page_t m)
401 {
402 
403 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
404 	if (bp->b_flags & B_CACHE) {
405 		int base = (foff + off) & PAGE_MASK;
406 		if (vm_page_is_valid(m, base, size) == 0)
407 			bp->b_flags &= ~B_CACHE;
408 	}
409 }
410 
411 /* Wake up the buffer deamon if necessary */
412 static __inline
413 void
414 bd_wakeup(int dirtybuflevel)
415 {
416 
417 	mtx_lock(&bdlock);
418 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
419 		bd_request = 1;
420 		wakeup(&bd_request);
421 	}
422 	mtx_unlock(&bdlock);
423 }
424 
425 /*
426  * bd_speedup - speedup the buffer cache flushing code
427  */
428 
429 static __inline
430 void
431 bd_speedup(void)
432 {
433 
434 	bd_wakeup(1);
435 }
436 
437 /*
438  * Calculating buffer cache scaling values and reserve space for buffer
439  * headers.  This is called during low level kernel initialization and
440  * may be called more then once.  We CANNOT write to the memory area
441  * being reserved at this time.
442  */
443 caddr_t
444 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
445 {
446 
447 	/*
448 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
449 	 * PAGE_SIZE is >= 1K)
450 	 */
451 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
452 
453 	/*
454 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
455 	 * For the first 64MB of ram nominally allocate sufficient buffers to
456 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
457 	 * buffers to cover 1/20 of our ram over 64MB.  When auto-sizing
458 	 * the buffer cache we limit the eventual kva reservation to
459 	 * maxbcache bytes.
460 	 *
461 	 * factor represents the 1/4 x ram conversion.
462 	 */
463 	if (nbuf == 0) {
464 		int factor = 4 * BKVASIZE / 1024;
465 
466 		nbuf = 50;
467 		if (physmem_est > 4096)
468 			nbuf += min((physmem_est - 4096) / factor,
469 			    65536 / factor);
470 		if (physmem_est > 65536)
471 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
472 
473 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
474 			nbuf = maxbcache / BKVASIZE;
475 	}
476 
477 #if 0
478 	/*
479 	 * Do not allow the buffer_map to be more then 1/2 the size of the
480 	 * kernel_map.
481 	 */
482 	if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
483 	    (BKVASIZE * 2)) {
484 		nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
485 		    (BKVASIZE * 2);
486 		printf("Warning: nbufs capped at %d\n", nbuf);
487 	}
488 #endif
489 
490 	/*
491 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
492 	 * We have no less then 16 and no more then 256.
493 	 */
494 	nswbuf = max(min(nbuf/4, 256), 16);
495 #ifdef NSWBUF_MIN
496 	if (nswbuf < NSWBUF_MIN)
497 		nswbuf = NSWBUF_MIN;
498 #endif
499 #ifdef DIRECTIO
500 	ffs_rawread_setup();
501 #endif
502 
503 	/*
504 	 * Reserve space for the buffer cache buffers
505 	 */
506 	swbuf = (void *)v;
507 	v = (caddr_t)(swbuf + nswbuf);
508 	buf = (void *)v;
509 	v = (caddr_t)(buf + nbuf);
510 
511 	return(v);
512 }
513 
514 /* Initialize the buffer subsystem.  Called before use of any buffers. */
515 void
516 bufinit(void)
517 {
518 	struct buf *bp;
519 	int i;
520 
521 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
522 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
523 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
524 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
525 	mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
526 
527 	/* next, make a null set of free lists */
528 	for (i = 0; i < BUFFER_QUEUES; i++)
529 		TAILQ_INIT(&bufqueues[i]);
530 
531 	/* finally, initialize each buffer header and stick on empty q */
532 	for (i = 0; i < nbuf; i++) {
533 		bp = &buf[i];
534 		bzero(bp, sizeof *bp);
535 		bp->b_flags = B_INVAL;	/* we're just an empty header */
536 		bp->b_rcred = NOCRED;
537 		bp->b_wcred = NOCRED;
538 		bp->b_qindex = QUEUE_EMPTY;
539 		bp->b_vflags = 0;
540 		bp->b_xflags = 0;
541 		LIST_INIT(&bp->b_dep);
542 		BUF_LOCKINIT(bp);
543 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
544 	}
545 
546 	/*
547 	 * maxbufspace is the absolute maximum amount of buffer space we are
548 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
549 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
550 	 * used by most other processes.  The differential is required to
551 	 * ensure that buf_daemon is able to run when other processes might
552 	 * be blocked waiting for buffer space.
553 	 *
554 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
555 	 * this may result in KVM fragmentation which is not handled optimally
556 	 * by the system.
557 	 */
558 	maxbufspace = nbuf * BKVASIZE;
559 	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
560 	lobufspace = hibufspace - MAXBSIZE;
561 
562 	lorunningspace = 512 * 1024;
563 	hirunningspace = 1024 * 1024;
564 
565 /*
566  * Limit the amount of malloc memory since it is wired permanently into
567  * the kernel space.  Even though this is accounted for in the buffer
568  * allocation, we don't want the malloced region to grow uncontrolled.
569  * The malloc scheme improves memory utilization significantly on average
570  * (small) directories.
571  */
572 	maxbufmallocspace = hibufspace / 20;
573 
574 /*
575  * Reduce the chance of a deadlock occuring by limiting the number
576  * of delayed-write dirty buffers we allow to stack up.
577  */
578 	hidirtybuffers = nbuf / 4 + 20;
579 	dirtybufthresh = hidirtybuffers * 9 / 10;
580 	numdirtybuffers = 0;
581 /*
582  * To support extreme low-memory systems, make sure hidirtybuffers cannot
583  * eat up all available buffer space.  This occurs when our minimum cannot
584  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
585  * BKVASIZE'd (8K) buffers.
586  */
587 	while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
588 		hidirtybuffers >>= 1;
589 	}
590 	lodirtybuffers = hidirtybuffers / 2;
591 
592 /*
593  * Try to keep the number of free buffers in the specified range,
594  * and give special processes (e.g. like buf_daemon) access to an
595  * emergency reserve.
596  */
597 	lofreebuffers = nbuf / 18 + 5;
598 	hifreebuffers = 2 * lofreebuffers;
599 	numfreebuffers = nbuf;
600 
601 /*
602  * Maximum number of async ops initiated per buf_daemon loop.  This is
603  * somewhat of a hack at the moment, we really need to limit ourselves
604  * based on the number of bytes of I/O in-transit that were initiated
605  * from buf_daemon.
606  */
607 
608 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
609 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
610 }
611 
612 /*
613  * bfreekva() - free the kva allocation for a buffer.
614  *
615  *	Must be called at splbio() or higher as this is the only locking for
616  *	buffer_map.
617  *
618  *	Since this call frees up buffer space, we call bufspacewakeup().
619  */
620 static void
621 bfreekva(struct buf *bp)
622 {
623 
624 	if (bp->b_kvasize) {
625 		atomic_add_int(&buffreekvacnt, 1);
626 		atomic_subtract_int(&bufspace, bp->b_kvasize);
627 		vm_map_delete(buffer_map,
628 		    (vm_offset_t) bp->b_kvabase,
629 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize
630 		);
631 		bp->b_kvasize = 0;
632 		bufspacewakeup();
633 	}
634 }
635 
636 /*
637  *	bremfree:
638  *
639  *	Mark the buffer for removal from the appropriate free list in brelse.
640  *
641  */
642 void
643 bremfree(struct buf *bp)
644 {
645 
646 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
647 	KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked."));
648 	KASSERT((bp->b_flags & B_REMFREE) == 0 && bp->b_qindex != QUEUE_NONE,
649 	    ("bremfree: buffer %p not on a queue.", bp));
650 
651 	bp->b_flags |= B_REMFREE;
652 	/* Fixup numfreebuffers count.  */
653 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
654 		atomic_subtract_int(&numfreebuffers, 1);
655 }
656 
657 /*
658  *	bremfreef:
659  *
660  *	Force an immediate removal from a free list.  Used only in nfs when
661  *	it abuses the b_freelist pointer.
662  */
663 void
664 bremfreef(struct buf *bp)
665 {
666 	mtx_lock(&bqlock);
667 	bremfreel(bp);
668 	mtx_unlock(&bqlock);
669 }
670 
671 /*
672  *	bremfreel:
673  *
674  *	Removes a buffer from the free list, must be called with the
675  *	bqlock held.
676  */
677 static void
678 bremfreel(struct buf *bp)
679 {
680 	int s = splbio();
681 
682 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
683 	    bp, bp->b_vp, bp->b_flags);
684 	KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp));
685 	KASSERT(bp->b_qindex != QUEUE_NONE,
686 	    ("bremfreel: buffer %p not on a queue.", bp));
687 	mtx_assert(&bqlock, MA_OWNED);
688 
689 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
690 	bp->b_qindex = QUEUE_NONE;
691 	/*
692 	 * If this was a delayed bremfree() we only need to remove the buffer
693 	 * from the queue and return the stats are already done.
694 	 */
695 	if (bp->b_flags & B_REMFREE) {
696 		bp->b_flags &= ~B_REMFREE;
697 		splx(s);
698 		return;
699 	}
700 	/*
701 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
702 	 * delayed-write, the buffer was free and we must decrement
703 	 * numfreebuffers.
704 	 */
705 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
706 		atomic_subtract_int(&numfreebuffers, 1);
707 	splx(s);
708 }
709 
710 
711 /*
712  * Get a buffer with the specified data.  Look in the cache first.  We
713  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
714  * is set, the buffer is valid and we do not have to do anything ( see
715  * getblk() ).  This is really just a special case of breadn().
716  */
717 int
718 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
719     struct buf **bpp)
720 {
721 
722 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
723 }
724 
725 /*
726  * Operates like bread, but also starts asynchronous I/O on
727  * read-ahead blocks.  We must clear BIO_ERROR and B_INVAL prior
728  * to initiating I/O . If B_CACHE is set, the buffer is valid
729  * and we do not have to do anything.
730  */
731 int
732 breadn(struct vnode * vp, daddr_t blkno, int size,
733     daddr_t * rablkno, int *rabsize,
734     int cnt, struct ucred * cred, struct buf **bpp)
735 {
736 	struct buf *bp, *rabp;
737 	int i;
738 	int rv = 0, readwait = 0;
739 
740 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
741 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
742 
743 	/* if not found in cache, do some I/O */
744 	if ((bp->b_flags & B_CACHE) == 0) {
745 		if (curthread != PCPU_GET(idlethread))
746 			curthread->td_proc->p_stats->p_ru.ru_inblock++;
747 		bp->b_iocmd = BIO_READ;
748 		bp->b_flags &= ~B_INVAL;
749 		bp->b_ioflags &= ~BIO_ERROR;
750 		if (bp->b_rcred == NOCRED && cred != NOCRED)
751 			bp->b_rcred = crhold(cred);
752 		vfs_busy_pages(bp, 0);
753 		bp->b_iooffset = dbtob(bp->b_blkno);
754 		bstrategy(bp);
755 		++readwait;
756 	}
757 
758 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
759 		if (inmem(vp, *rablkno))
760 			continue;
761 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
762 
763 		if ((rabp->b_flags & B_CACHE) == 0) {
764 			if (curthread != PCPU_GET(idlethread))
765 				curthread->td_proc->p_stats->p_ru.ru_inblock++;
766 			rabp->b_flags |= B_ASYNC;
767 			rabp->b_flags &= ~B_INVAL;
768 			rabp->b_ioflags &= ~BIO_ERROR;
769 			rabp->b_iocmd = BIO_READ;
770 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
771 				rabp->b_rcred = crhold(cred);
772 			vfs_busy_pages(rabp, 0);
773 			BUF_KERNPROC(rabp);
774 			rabp->b_iooffset = dbtob(rabp->b_blkno);
775 			bstrategy(rabp);
776 		} else {
777 			brelse(rabp);
778 		}
779 	}
780 
781 	if (readwait) {
782 		rv = bufwait(bp);
783 	}
784 	return (rv);
785 }
786 
787 /*
788  * Write, release buffer on completion.  (Done by iodone
789  * if async).  Do not bother writing anything if the buffer
790  * is invalid.
791  *
792  * Note that we set B_CACHE here, indicating that buffer is
793  * fully valid and thus cacheable.  This is true even of NFS
794  * now so we set it generally.  This could be set either here
795  * or in biodone() since the I/O is synchronous.  We put it
796  * here.
797  */
798 int
799 bufwrite(struct buf *bp)
800 {
801 	int oldflags, s;
802 
803 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
804 	if (bp->b_flags & B_INVAL) {
805 		brelse(bp);
806 		return (0);
807 	}
808 
809 	oldflags = bp->b_flags;
810 
811 	if (BUF_REFCNT(bp) == 0)
812 		panic("bufwrite: buffer is not busy???");
813 	s = splbio();
814 
815 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
816 	    ("FFS background buffer should not get here %p", bp));
817 
818 	/* Mark the buffer clean */
819 	bundirty(bp);
820 
821 	bp->b_flags &= ~B_DONE;
822 	bp->b_ioflags &= ~BIO_ERROR;
823 	bp->b_flags |= B_CACHE;
824 	bp->b_iocmd = BIO_WRITE;
825 
826 	bufobj_wref(bp->b_bufobj);
827 	vfs_busy_pages(bp, 1);
828 
829 	/*
830 	 * Normal bwrites pipeline writes
831 	 */
832 	bp->b_runningbufspace = bp->b_bufsize;
833 	atomic_add_int(&runningbufspace, bp->b_runningbufspace);
834 
835 	if (curthread != PCPU_GET(idlethread))
836 		curthread->td_proc->p_stats->p_ru.ru_oublock++;
837 	splx(s);
838 	if (oldflags & B_ASYNC)
839 		BUF_KERNPROC(bp);
840 	bp->b_iooffset = dbtob(bp->b_blkno);
841 	bstrategy(bp);
842 
843 	if ((oldflags & B_ASYNC) == 0) {
844 		int rtval = bufwait(bp);
845 		brelse(bp);
846 		return (rtval);
847 	} else {
848 		/*
849 		 * don't allow the async write to saturate the I/O
850 		 * system.  We will not deadlock here because
851 		 * we are blocking waiting for I/O that is already in-progress
852 		 * to complete. We do not block here if it is the update
853 		 * or syncer daemon trying to clean up as that can lead
854 		 * to deadlock.
855 		 */
856 		if (curthread->td_proc != bufdaemonproc &&
857 		    curthread->td_proc != updateproc)
858 			waitrunningbufspace();
859 	}
860 
861 	return (0);
862 }
863 
864 /*
865  * Delayed write. (Buffer is marked dirty).  Do not bother writing
866  * anything if the buffer is marked invalid.
867  *
868  * Note that since the buffer must be completely valid, we can safely
869  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
870  * biodone() in order to prevent getblk from writing the buffer
871  * out synchronously.
872  */
873 void
874 bdwrite(struct buf *bp)
875 {
876 	struct thread *td = curthread;
877 	struct vnode *vp;
878 	struct buf *nbp;
879 	struct bufobj *bo;
880 
881 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
882 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
883 	KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy"));
884 
885 	if (bp->b_flags & B_INVAL) {
886 		brelse(bp);
887 		return;
888 	}
889 
890 	/*
891 	 * If we have too many dirty buffers, don't create any more.
892 	 * If we are wildly over our limit, then force a complete
893 	 * cleanup. Otherwise, just keep the situation from getting
894 	 * out of control. Note that we have to avoid a recursive
895 	 * disaster and not try to clean up after our own cleanup!
896 	 */
897 	vp = bp->b_vp;
898 	bo = bp->b_bufobj;
899 	if ((td->td_pflags & TDP_COWINPROGRESS) == 0) {
900 		BO_LOCK(bo);
901 		if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
902 			BO_UNLOCK(bo);
903 			(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
904 			altbufferflushes++;
905 		} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
906 			/*
907 			 * Try to find a buffer to flush.
908 			 */
909 			TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
910 				if ((nbp->b_vflags & BV_BKGRDINPROG) ||
911 				    BUF_LOCK(nbp,
912 				    LK_EXCLUSIVE | LK_NOWAIT, NULL))
913 					continue;
914 				if (bp == nbp)
915 					panic("bdwrite: found ourselves");
916 				BO_UNLOCK(bo);
917 				/* Don't countdeps with the bo lock held. */
918 				if (buf_countdeps(nbp, 0)) {
919 					BO_LOCK(bo);
920 					BUF_UNLOCK(nbp);
921 					continue;
922 				}
923 				if (nbp->b_flags & B_CLUSTEROK) {
924 					vfs_bio_awrite(nbp);
925 				} else {
926 					bremfree(nbp);
927 					bawrite(nbp);
928 				}
929 				dirtybufferflushes++;
930 				break;
931 			}
932 			if (nbp == NULL)
933 				BO_UNLOCK(bo);
934 		} else
935 			BO_UNLOCK(bo);
936 	} else
937 		recursiveflushes++;
938 
939 	bdirty(bp);
940 	/*
941 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
942 	 * true even of NFS now.
943 	 */
944 	bp->b_flags |= B_CACHE;
945 
946 	/*
947 	 * This bmap keeps the system from needing to do the bmap later,
948 	 * perhaps when the system is attempting to do a sync.  Since it
949 	 * is likely that the indirect block -- or whatever other datastructure
950 	 * that the filesystem needs is still in memory now, it is a good
951 	 * thing to do this.  Note also, that if the pageout daemon is
952 	 * requesting a sync -- there might not be enough memory to do
953 	 * the bmap then...  So, this is important to do.
954 	 */
955 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
956 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
957 	}
958 
959 	/*
960 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
961 	 */
962 	vfs_setdirty(bp);
963 
964 	/*
965 	 * We need to do this here to satisfy the vnode_pager and the
966 	 * pageout daemon, so that it thinks that the pages have been
967 	 * "cleaned".  Note that since the pages are in a delayed write
968 	 * buffer -- the VFS layer "will" see that the pages get written
969 	 * out on the next sync, or perhaps the cluster will be completed.
970 	 */
971 	vfs_clean_pages(bp);
972 	bqrelse(bp);
973 
974 	/*
975 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
976 	 * buffers (midpoint between our recovery point and our stall
977 	 * point).
978 	 */
979 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
980 
981 	/*
982 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
983 	 * due to the softdep code.
984 	 */
985 }
986 
987 /*
988  *	bdirty:
989  *
990  *	Turn buffer into delayed write request.  We must clear BIO_READ and
991  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
992  *	itself to properly update it in the dirty/clean lists.  We mark it
993  *	B_DONE to ensure that any asynchronization of the buffer properly
994  *	clears B_DONE ( else a panic will occur later ).
995  *
996  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
997  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
998  *	should only be called if the buffer is known-good.
999  *
1000  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1001  *	count.
1002  *
1003  *	Must be called at splbio().
1004  *	The buffer must be on QUEUE_NONE.
1005  */
1006 void
1007 bdirty(struct buf *bp)
1008 {
1009 
1010 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1011 	    bp, bp->b_vp, bp->b_flags);
1012 	KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp));
1013 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1014 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1015 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1016 	bp->b_flags &= ~(B_RELBUF);
1017 	bp->b_iocmd = BIO_WRITE;
1018 
1019 	if ((bp->b_flags & B_DELWRI) == 0) {
1020 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1021 		reassignbuf(bp);
1022 		atomic_add_int(&numdirtybuffers, 1);
1023 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1024 	}
1025 }
1026 
1027 /*
1028  *	bundirty:
1029  *
1030  *	Clear B_DELWRI for buffer.
1031  *
1032  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1033  *	count.
1034  *
1035  *	Must be called at splbio().
1036  *	The buffer must be on QUEUE_NONE.
1037  */
1038 
1039 void
1040 bundirty(struct buf *bp)
1041 {
1042 
1043 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1044 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1045 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1046 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1047 	KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp));
1048 
1049 	if (bp->b_flags & B_DELWRI) {
1050 		bp->b_flags &= ~B_DELWRI;
1051 		reassignbuf(bp);
1052 		atomic_subtract_int(&numdirtybuffers, 1);
1053 		numdirtywakeup(lodirtybuffers);
1054 	}
1055 	/*
1056 	 * Since it is now being written, we can clear its deferred write flag.
1057 	 */
1058 	bp->b_flags &= ~B_DEFERRED;
1059 }
1060 
1061 /*
1062  *	bawrite:
1063  *
1064  *	Asynchronous write.  Start output on a buffer, but do not wait for
1065  *	it to complete.  The buffer is released when the output completes.
1066  *
1067  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1068  *	B_INVAL buffers.  Not us.
1069  */
1070 void
1071 bawrite(struct buf *bp)
1072 {
1073 
1074 	bp->b_flags |= B_ASYNC;
1075 	(void) bwrite(bp);
1076 }
1077 
1078 /*
1079  *	bwillwrite:
1080  *
1081  *	Called prior to the locking of any vnodes when we are expecting to
1082  *	write.  We do not want to starve the buffer cache with too many
1083  *	dirty buffers so we block here.  By blocking prior to the locking
1084  *	of any vnodes we attempt to avoid the situation where a locked vnode
1085  *	prevents the various system daemons from flushing related buffers.
1086  */
1087 
1088 void
1089 bwillwrite(void)
1090 {
1091 
1092 	if (numdirtybuffers >= hidirtybuffers) {
1093 		int s;
1094 
1095 		s = splbio();
1096 		mtx_lock(&nblock);
1097 		while (numdirtybuffers >= hidirtybuffers) {
1098 			bd_wakeup(1);
1099 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1100 			msleep(&needsbuffer, &nblock,
1101 			    (PRIBIO + 4), "flswai", 0);
1102 		}
1103 		splx(s);
1104 		mtx_unlock(&nblock);
1105 	}
1106 }
1107 
1108 /*
1109  * Return true if we have too many dirty buffers.
1110  */
1111 int
1112 buf_dirty_count_severe(void)
1113 {
1114 
1115 	return(numdirtybuffers >= hidirtybuffers);
1116 }
1117 
1118 /*
1119  *	brelse:
1120  *
1121  *	Release a busy buffer and, if requested, free its resources.  The
1122  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1123  *	to be accessed later as a cache entity or reused for other purposes.
1124  */
1125 void
1126 brelse(struct buf *bp)
1127 {
1128 	int s;
1129 
1130 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1131 	    bp, bp->b_vp, bp->b_flags);
1132 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1133 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1134 
1135 	s = splbio();
1136 
1137 	if (bp->b_iocmd == BIO_WRITE &&
1138 	    (bp->b_ioflags & BIO_ERROR) &&
1139 	    !(bp->b_flags & B_INVAL)) {
1140 		/*
1141 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1142 		 * pages from being scrapped.  If B_INVAL is set then
1143 		 * this case is not run and the next case is run to
1144 		 * destroy the buffer.  B_INVAL can occur if the buffer
1145 		 * is outside the range supported by the underlying device.
1146 		 */
1147 		bp->b_ioflags &= ~BIO_ERROR;
1148 		bdirty(bp);
1149 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1150 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1151 		/*
1152 		 * Either a failed I/O or we were asked to free or not
1153 		 * cache the buffer.
1154 		 */
1155 		bp->b_flags |= B_INVAL;
1156 		if (LIST_FIRST(&bp->b_dep) != NULL)
1157 			buf_deallocate(bp);
1158 		if (bp->b_flags & B_DELWRI) {
1159 			atomic_subtract_int(&numdirtybuffers, 1);
1160 			numdirtywakeup(lodirtybuffers);
1161 		}
1162 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1163 		if ((bp->b_flags & B_VMIO) == 0) {
1164 			if (bp->b_bufsize)
1165 				allocbuf(bp, 0);
1166 			if (bp->b_vp)
1167 				brelvp(bp);
1168 		}
1169 	}
1170 
1171 	/*
1172 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1173 	 * is called with B_DELWRI set, the underlying pages may wind up
1174 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1175 	 * because pages associated with a B_DELWRI bp are marked clean.
1176 	 *
1177 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1178 	 * if B_DELWRI is set.
1179 	 *
1180 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1181 	 * on pages to return pages to the VM page queues.
1182 	 */
1183 	if (bp->b_flags & B_DELWRI)
1184 		bp->b_flags &= ~B_RELBUF;
1185 	else if (vm_page_count_severe()) {
1186 		/*
1187 		 * XXX This lock may not be necessary since BKGRDINPROG
1188 		 * cannot be set while we hold the buf lock, it can only be
1189 		 * cleared if it is already pending.
1190 		 */
1191 		if (bp->b_vp) {
1192 			BO_LOCK(bp->b_bufobj);
1193 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1194 				bp->b_flags |= B_RELBUF;
1195 			BO_UNLOCK(bp->b_bufobj);
1196 		} else
1197 			bp->b_flags |= B_RELBUF;
1198 	}
1199 
1200 	/*
1201 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1202 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1203 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1204 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1205 	 *
1206 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1207 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1208 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1209 	 *
1210 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1211 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1212 	 * the commit state and we cannot afford to lose the buffer. If the
1213 	 * buffer has a background write in progress, we need to keep it
1214 	 * around to prevent it from being reconstituted and starting a second
1215 	 * background write.
1216 	 */
1217 	if ((bp->b_flags & B_VMIO)
1218 	    && !(bp->b_vp->v_mount != NULL &&
1219 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1220 		 !vn_isdisk(bp->b_vp, NULL) &&
1221 		 (bp->b_flags & B_DELWRI))
1222 	    ) {
1223 
1224 		int i, j, resid;
1225 		vm_page_t m;
1226 		off_t foff;
1227 		vm_pindex_t poff;
1228 		vm_object_t obj;
1229 
1230 		obj = bp->b_bufobj->bo_object;
1231 
1232 		/*
1233 		 * Get the base offset and length of the buffer.  Note that
1234 		 * in the VMIO case if the buffer block size is not
1235 		 * page-aligned then b_data pointer may not be page-aligned.
1236 		 * But our b_pages[] array *IS* page aligned.
1237 		 *
1238 		 * block sizes less then DEV_BSIZE (usually 512) are not
1239 		 * supported due to the page granularity bits (m->valid,
1240 		 * m->dirty, etc...).
1241 		 *
1242 		 * See man buf(9) for more information
1243 		 */
1244 		resid = bp->b_bufsize;
1245 		foff = bp->b_offset;
1246 		VM_OBJECT_LOCK(obj);
1247 		for (i = 0; i < bp->b_npages; i++) {
1248 			int had_bogus = 0;
1249 
1250 			m = bp->b_pages[i];
1251 
1252 			/*
1253 			 * If we hit a bogus page, fixup *all* the bogus pages
1254 			 * now.
1255 			 */
1256 			if (m == bogus_page) {
1257 				poff = OFF_TO_IDX(bp->b_offset);
1258 				had_bogus = 1;
1259 
1260 				for (j = i; j < bp->b_npages; j++) {
1261 					vm_page_t mtmp;
1262 					mtmp = bp->b_pages[j];
1263 					if (mtmp == bogus_page) {
1264 						mtmp = vm_page_lookup(obj, poff + j);
1265 						if (!mtmp) {
1266 							panic("brelse: page missing\n");
1267 						}
1268 						bp->b_pages[j] = mtmp;
1269 					}
1270 				}
1271 
1272 				if ((bp->b_flags & B_INVAL) == 0) {
1273 					pmap_qenter(
1274 					    trunc_page((vm_offset_t)bp->b_data),
1275 					    bp->b_pages, bp->b_npages);
1276 				}
1277 				m = bp->b_pages[i];
1278 			}
1279 			if ((bp->b_flags & B_NOCACHE) ||
1280 			    (bp->b_ioflags & BIO_ERROR)) {
1281 				int poffset = foff & PAGE_MASK;
1282 				int presid = resid > (PAGE_SIZE - poffset) ?
1283 					(PAGE_SIZE - poffset) : resid;
1284 
1285 				KASSERT(presid >= 0, ("brelse: extra page"));
1286 				vm_page_lock_queues();
1287 				vm_page_set_invalid(m, poffset, presid);
1288 				vm_page_unlock_queues();
1289 				if (had_bogus)
1290 					printf("avoided corruption bug in bogus_page/brelse code\n");
1291 			}
1292 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1293 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1294 		}
1295 		VM_OBJECT_UNLOCK(obj);
1296 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1297 			vfs_vmio_release(bp);
1298 
1299 	} else if (bp->b_flags & B_VMIO) {
1300 
1301 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1302 			vfs_vmio_release(bp);
1303 		}
1304 
1305 	}
1306 
1307 	if (BUF_REFCNT(bp) > 1) {
1308 		/* do not release to free list */
1309 		BUF_UNLOCK(bp);
1310 		splx(s);
1311 		return;
1312 	}
1313 
1314 	/* enqueue */
1315 	mtx_lock(&bqlock);
1316 	/* Handle delayed bremfree() processing. */
1317 	if (bp->b_flags & B_REMFREE)
1318 		bremfreel(bp);
1319 	if (bp->b_qindex != QUEUE_NONE)
1320 		panic("brelse: free buffer onto another queue???");
1321 
1322 	/* buffers with no memory */
1323 	if (bp->b_bufsize == 0) {
1324 		bp->b_flags |= B_INVAL;
1325 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1326 		if (bp->b_vflags & BV_BKGRDINPROG)
1327 			panic("losing buffer 1");
1328 		if (bp->b_kvasize) {
1329 			bp->b_qindex = QUEUE_EMPTYKVA;
1330 		} else {
1331 			bp->b_qindex = QUEUE_EMPTY;
1332 		}
1333 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1334 	/* buffers with junk contents */
1335 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1336 	    (bp->b_ioflags & BIO_ERROR)) {
1337 		bp->b_flags |= B_INVAL;
1338 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1339 		if (bp->b_vflags & BV_BKGRDINPROG)
1340 			panic("losing buffer 2");
1341 		bp->b_qindex = QUEUE_CLEAN;
1342 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1343 	/* remaining buffers */
1344 	} else {
1345 		if (bp->b_flags & B_DELWRI)
1346 			bp->b_qindex = QUEUE_DIRTY;
1347 		else
1348 			bp->b_qindex = QUEUE_CLEAN;
1349 		if (bp->b_flags & B_AGE)
1350 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1351 		else
1352 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1353 	}
1354 	mtx_unlock(&bqlock);
1355 
1356 	/*
1357 	 * If B_INVAL and B_DELWRI is set, clear B_DELWRI.  We have already
1358 	 * placed the buffer on the correct queue.  We must also disassociate
1359 	 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1360 	 * find it.
1361 	 */
1362 	if (bp->b_flags & B_INVAL) {
1363 		if (bp->b_flags & B_DELWRI)
1364 			bundirty(bp);
1365 		if (bp->b_vp)
1366 			brelvp(bp);
1367 	}
1368 
1369 	/*
1370 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1371 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1372 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1373 	 * if B_INVAL is set ).
1374 	 */
1375 
1376 	if (!(bp->b_flags & B_DELWRI))
1377 		bufcountwakeup();
1378 
1379 	/*
1380 	 * Something we can maybe free or reuse
1381 	 */
1382 	if (bp->b_bufsize || bp->b_kvasize)
1383 		bufspacewakeup();
1384 
1385 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1386 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1387 		panic("brelse: not dirty");
1388 	/* unlock */
1389 	BUF_UNLOCK(bp);
1390 	splx(s);
1391 }
1392 
1393 /*
1394  * Release a buffer back to the appropriate queue but do not try to free
1395  * it.  The buffer is expected to be used again soon.
1396  *
1397  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1398  * biodone() to requeue an async I/O on completion.  It is also used when
1399  * known good buffers need to be requeued but we think we may need the data
1400  * again soon.
1401  *
1402  * XXX we should be able to leave the B_RELBUF hint set on completion.
1403  */
1404 void
1405 bqrelse(struct buf *bp)
1406 {
1407 	int s;
1408 
1409 	s = splbio();
1410 
1411 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1412 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1413 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1414 
1415 	if (BUF_REFCNT(bp) > 1) {
1416 		/* do not release to free list */
1417 		BUF_UNLOCK(bp);
1418 		splx(s);
1419 		return;
1420 	}
1421 	mtx_lock(&bqlock);
1422 	/* Handle delayed bremfree() processing. */
1423 	if (bp->b_flags & B_REMFREE)
1424 		bremfreel(bp);
1425 	if (bp->b_qindex != QUEUE_NONE)
1426 		panic("bqrelse: free buffer onto another queue???");
1427 	/* buffers with stale but valid contents */
1428 	if (bp->b_flags & B_DELWRI) {
1429 		bp->b_qindex = QUEUE_DIRTY;
1430 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1431 	} else {
1432 		/*
1433 		 * XXX This lock may not be necessary since BKGRDINPROG
1434 		 * cannot be set while we hold the buf lock, it can only be
1435 		 * cleared if it is already pending.
1436 		 */
1437 		BO_LOCK(bp->b_bufobj);
1438 		if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1439 			BO_UNLOCK(bp->b_bufobj);
1440 			bp->b_qindex = QUEUE_CLEAN;
1441 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1442 			    b_freelist);
1443 		} else {
1444 			/*
1445 			 * We are too low on memory, we have to try to free
1446 			 * the buffer (most importantly: the wired pages
1447 			 * making up its backing store) *now*.
1448 			 */
1449 			BO_UNLOCK(bp->b_bufobj);
1450 			mtx_unlock(&bqlock);
1451 			splx(s);
1452 			brelse(bp);
1453 			return;
1454 		}
1455 	}
1456 	mtx_unlock(&bqlock);
1457 
1458 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1459 		bufcountwakeup();
1460 
1461 	/*
1462 	 * Something we can maybe free or reuse.
1463 	 */
1464 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1465 		bufspacewakeup();
1466 
1467 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1468 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1469 		panic("bqrelse: not dirty");
1470 	/* unlock */
1471 	BUF_UNLOCK(bp);
1472 	splx(s);
1473 }
1474 
1475 /* Give pages used by the bp back to the VM system (where possible) */
1476 static void
1477 vfs_vmio_release(struct buf *bp)
1478 {
1479 	int i;
1480 	vm_page_t m;
1481 
1482 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1483 	vm_page_lock_queues();
1484 	for (i = 0; i < bp->b_npages; i++) {
1485 		m = bp->b_pages[i];
1486 		bp->b_pages[i] = NULL;
1487 		/*
1488 		 * In order to keep page LRU ordering consistent, put
1489 		 * everything on the inactive queue.
1490 		 */
1491 		vm_page_unwire(m, 0);
1492 		/*
1493 		 * We don't mess with busy pages, it is
1494 		 * the responsibility of the process that
1495 		 * busied the pages to deal with them.
1496 		 */
1497 		if ((m->flags & PG_BUSY) || (m->busy != 0))
1498 			continue;
1499 
1500 		if (m->wire_count == 0) {
1501 			/*
1502 			 * Might as well free the page if we can and it has
1503 			 * no valid data.  We also free the page if the
1504 			 * buffer was used for direct I/O
1505 			 */
1506 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1507 			    m->hold_count == 0) {
1508 				pmap_remove_all(m);
1509 				vm_page_free(m);
1510 			} else if (bp->b_flags & B_DIRECT) {
1511 				vm_page_try_to_free(m);
1512 			} else if (vm_page_count_severe()) {
1513 				vm_page_try_to_cache(m);
1514 			}
1515 		}
1516 	}
1517 	vm_page_unlock_queues();
1518 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1519 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1520 
1521 	if (bp->b_bufsize) {
1522 		bufspacewakeup();
1523 		bp->b_bufsize = 0;
1524 	}
1525 	bp->b_npages = 0;
1526 	bp->b_flags &= ~B_VMIO;
1527 	if (bp->b_vp)
1528 		brelvp(bp);
1529 }
1530 
1531 /*
1532  * Check to see if a block at a particular lbn is available for a clustered
1533  * write.
1534  */
1535 static int
1536 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1537 {
1538 	struct buf *bpa;
1539 	int match;
1540 
1541 	match = 0;
1542 
1543 	/* If the buf isn't in core skip it */
1544 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1545 		return (0);
1546 
1547 	/* If the buf is busy we don't want to wait for it */
1548 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1549 		return (0);
1550 
1551 	/* Only cluster with valid clusterable delayed write buffers */
1552 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1553 	    (B_DELWRI | B_CLUSTEROK))
1554 		goto done;
1555 
1556 	if (bpa->b_bufsize != size)
1557 		goto done;
1558 
1559 	/*
1560 	 * Check to see if it is in the expected place on disk and that the
1561 	 * block has been mapped.
1562 	 */
1563 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1564 		match = 1;
1565 done:
1566 	BUF_UNLOCK(bpa);
1567 	return (match);
1568 }
1569 
1570 /*
1571  *	vfs_bio_awrite:
1572  *
1573  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1574  *	This is much better then the old way of writing only one buffer at
1575  *	a time.  Note that we may not be presented with the buffers in the
1576  *	correct order, so we search for the cluster in both directions.
1577  */
1578 int
1579 vfs_bio_awrite(struct buf *bp)
1580 {
1581 	int i;
1582 	int j;
1583 	daddr_t lblkno = bp->b_lblkno;
1584 	struct vnode *vp = bp->b_vp;
1585 	int s;
1586 	int ncl;
1587 	int nwritten;
1588 	int size;
1589 	int maxcl;
1590 
1591 	s = splbio();
1592 	/*
1593 	 * right now we support clustered writing only to regular files.  If
1594 	 * we find a clusterable block we could be in the middle of a cluster
1595 	 * rather then at the beginning.
1596 	 */
1597 	if ((vp->v_type == VREG) &&
1598 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1599 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1600 
1601 		size = vp->v_mount->mnt_stat.f_iosize;
1602 		maxcl = MAXPHYS / size;
1603 
1604 		VI_LOCK(vp);
1605 		for (i = 1; i < maxcl; i++)
1606 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1607 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1608 				break;
1609 
1610 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1611 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1612 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1613 				break;
1614 
1615 		VI_UNLOCK(vp);
1616 		--j;
1617 		ncl = i + j;
1618 		/*
1619 		 * this is a possible cluster write
1620 		 */
1621 		if (ncl != 1) {
1622 			BUF_UNLOCK(bp);
1623 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1624 			splx(s);
1625 			return nwritten;
1626 		}
1627 	}
1628 
1629 	bremfree(bp);
1630 	bp->b_flags |= B_ASYNC;
1631 
1632 	splx(s);
1633 	/*
1634 	 * default (old) behavior, writing out only one block
1635 	 *
1636 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1637 	 */
1638 	nwritten = bp->b_bufsize;
1639 	(void) bwrite(bp);
1640 
1641 	return nwritten;
1642 }
1643 
1644 /*
1645  *	getnewbuf:
1646  *
1647  *	Find and initialize a new buffer header, freeing up existing buffers
1648  *	in the bufqueues as necessary.  The new buffer is returned locked.
1649  *
1650  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1651  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1652  *
1653  *	We block if:
1654  *		We have insufficient buffer headers
1655  *		We have insufficient buffer space
1656  *		buffer_map is too fragmented ( space reservation fails )
1657  *		If we have to flush dirty buffers ( but we try to avoid this )
1658  *
1659  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1660  *	Instead we ask the buf daemon to do it for us.  We attempt to
1661  *	avoid piecemeal wakeups of the pageout daemon.
1662  */
1663 
1664 static struct buf *
1665 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1666 {
1667 	struct buf *bp;
1668 	struct buf *nbp;
1669 	int defrag = 0;
1670 	int nqindex;
1671 	static int flushingbufs;
1672 
1673 	/*
1674 	 * We can't afford to block since we might be holding a vnode lock,
1675 	 * which may prevent system daemons from running.  We deal with
1676 	 * low-memory situations by proactively returning memory and running
1677 	 * async I/O rather then sync I/O.
1678 	 */
1679 
1680 	atomic_add_int(&getnewbufcalls, 1);
1681 	atomic_subtract_int(&getnewbufrestarts, 1);
1682 restart:
1683 	atomic_add_int(&getnewbufrestarts, 1);
1684 
1685 	/*
1686 	 * Setup for scan.  If we do not have enough free buffers,
1687 	 * we setup a degenerate case that immediately fails.  Note
1688 	 * that if we are specially marked process, we are allowed to
1689 	 * dip into our reserves.
1690 	 *
1691 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1692 	 *
1693 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1694 	 * However, there are a number of cases (defragging, reusing, ...)
1695 	 * where we cannot backup.
1696 	 */
1697 	mtx_lock(&bqlock);
1698 	nqindex = QUEUE_EMPTYKVA;
1699 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1700 
1701 	if (nbp == NULL) {
1702 		/*
1703 		 * If no EMPTYKVA buffers and we are either
1704 		 * defragging or reusing, locate a CLEAN buffer
1705 		 * to free or reuse.  If bufspace useage is low
1706 		 * skip this step so we can allocate a new buffer.
1707 		 */
1708 		if (defrag || bufspace >= lobufspace) {
1709 			nqindex = QUEUE_CLEAN;
1710 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1711 		}
1712 
1713 		/*
1714 		 * If we could not find or were not allowed to reuse a
1715 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1716 		 * buffer.  We can only use an EMPTY buffer if allocating
1717 		 * its KVA would not otherwise run us out of buffer space.
1718 		 */
1719 		if (nbp == NULL && defrag == 0 &&
1720 		    bufspace + maxsize < hibufspace) {
1721 			nqindex = QUEUE_EMPTY;
1722 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1723 		}
1724 	}
1725 
1726 	/*
1727 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1728 	 * depending.
1729 	 */
1730 
1731 	while ((bp = nbp) != NULL) {
1732 		int qindex = nqindex;
1733 
1734 		/*
1735 		 * Calculate next bp ( we can only use it if we do not block
1736 		 * or do other fancy things ).
1737 		 */
1738 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1739 			switch(qindex) {
1740 			case QUEUE_EMPTY:
1741 				nqindex = QUEUE_EMPTYKVA;
1742 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1743 					break;
1744 				/* FALLTHROUGH */
1745 			case QUEUE_EMPTYKVA:
1746 				nqindex = QUEUE_CLEAN;
1747 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1748 					break;
1749 				/* FALLTHROUGH */
1750 			case QUEUE_CLEAN:
1751 				/*
1752 				 * nbp is NULL.
1753 				 */
1754 				break;
1755 			}
1756 		}
1757 		/*
1758 		 * If we are defragging then we need a buffer with
1759 		 * b_kvasize != 0.  XXX this situation should no longer
1760 		 * occur, if defrag is non-zero the buffer's b_kvasize
1761 		 * should also be non-zero at this point.  XXX
1762 		 */
1763 		if (defrag && bp->b_kvasize == 0) {
1764 			printf("Warning: defrag empty buffer %p\n", bp);
1765 			continue;
1766 		}
1767 
1768 		/*
1769 		 * Start freeing the bp.  This is somewhat involved.  nbp
1770 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1771 		 */
1772 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1773 			continue;
1774 		if (bp->b_vp) {
1775 			BO_LOCK(bp->b_bufobj);
1776 			if (bp->b_vflags & BV_BKGRDINPROG) {
1777 				BO_UNLOCK(bp->b_bufobj);
1778 				BUF_UNLOCK(bp);
1779 				continue;
1780 			}
1781 			BO_UNLOCK(bp->b_bufobj);
1782 		}
1783 		CTR6(KTR_BUF,
1784 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1785 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1786 		    bp->b_kvasize, bp->b_bufsize, qindex);
1787 
1788 		/*
1789 		 * Sanity Checks
1790 		 */
1791 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1792 
1793 		/*
1794 		 * Note: we no longer distinguish between VMIO and non-VMIO
1795 		 * buffers.
1796 		 */
1797 
1798 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1799 
1800 		bremfreel(bp);
1801 		mtx_unlock(&bqlock);
1802 
1803 		if (qindex == QUEUE_CLEAN) {
1804 			if (bp->b_flags & B_VMIO) {
1805 				bp->b_flags &= ~B_ASYNC;
1806 				vfs_vmio_release(bp);
1807 			}
1808 			if (bp->b_vp)
1809 				brelvp(bp);
1810 		}
1811 
1812 		/*
1813 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1814 		 * the scan from this point on.
1815 		 *
1816 		 * Get the rest of the buffer freed up.  b_kva* is still
1817 		 * valid after this operation.
1818 		 */
1819 
1820 		if (bp->b_rcred != NOCRED) {
1821 			crfree(bp->b_rcred);
1822 			bp->b_rcred = NOCRED;
1823 		}
1824 		if (bp->b_wcred != NOCRED) {
1825 			crfree(bp->b_wcred);
1826 			bp->b_wcred = NOCRED;
1827 		}
1828 		if (LIST_FIRST(&bp->b_dep) != NULL)
1829 			buf_deallocate(bp);
1830 		if (bp->b_vflags & BV_BKGRDINPROG)
1831 			panic("losing buffer 3");
1832 
1833 		if (bp->b_bufsize)
1834 			allocbuf(bp, 0);
1835 
1836 		bp->b_flags = 0;
1837 		bp->b_ioflags = 0;
1838 		bp->b_xflags = 0;
1839 		bp->b_vflags = 0;
1840 		bp->b_vp = NULL;
1841 		bp->b_blkno = bp->b_lblkno = 0;
1842 		bp->b_offset = NOOFFSET;
1843 		bp->b_iodone = 0;
1844 		bp->b_error = 0;
1845 		bp->b_resid = 0;
1846 		bp->b_bcount = 0;
1847 		bp->b_npages = 0;
1848 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1849 		bp->b_bufobj = NULL;
1850 
1851 		LIST_INIT(&bp->b_dep);
1852 
1853 		/*
1854 		 * If we are defragging then free the buffer.
1855 		 */
1856 		if (defrag) {
1857 			bp->b_flags |= B_INVAL;
1858 			bfreekva(bp);
1859 			brelse(bp);
1860 			defrag = 0;
1861 			goto restart;
1862 		}
1863 
1864 		/*
1865 		 * If we are overcomitted then recover the buffer and its
1866 		 * KVM space.  This occurs in rare situations when multiple
1867 		 * processes are blocked in getnewbuf() or allocbuf().
1868 		 */
1869 		if (bufspace >= hibufspace)
1870 			flushingbufs = 1;
1871 		if (flushingbufs && bp->b_kvasize != 0) {
1872 			bp->b_flags |= B_INVAL;
1873 			bfreekva(bp);
1874 			brelse(bp);
1875 			goto restart;
1876 		}
1877 		if (bufspace < lobufspace)
1878 			flushingbufs = 0;
1879 		break;
1880 	}
1881 
1882 	/*
1883 	 * If we exhausted our list, sleep as appropriate.  We may have to
1884 	 * wakeup various daemons and write out some dirty buffers.
1885 	 *
1886 	 * Generally we are sleeping due to insufficient buffer space.
1887 	 */
1888 
1889 	if (bp == NULL) {
1890 		int flags;
1891 		char *waitmsg;
1892 
1893 		mtx_unlock(&bqlock);
1894 		if (defrag) {
1895 			flags = VFS_BIO_NEED_BUFSPACE;
1896 			waitmsg = "nbufkv";
1897 		} else if (bufspace >= hibufspace) {
1898 			waitmsg = "nbufbs";
1899 			flags = VFS_BIO_NEED_BUFSPACE;
1900 		} else {
1901 			waitmsg = "newbuf";
1902 			flags = VFS_BIO_NEED_ANY;
1903 		}
1904 
1905 		bd_speedup();	/* heeeelp */
1906 
1907 		mtx_lock(&nblock);
1908 		needsbuffer |= flags;
1909 		while (needsbuffer & flags) {
1910 			if (msleep(&needsbuffer, &nblock,
1911 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1912 				mtx_unlock(&nblock);
1913 				return (NULL);
1914 			}
1915 		}
1916 		mtx_unlock(&nblock);
1917 	} else {
1918 		/*
1919 		 * We finally have a valid bp.  We aren't quite out of the
1920 		 * woods, we still have to reserve kva space.  In order
1921 		 * to keep fragmentation sane we only allocate kva in
1922 		 * BKVASIZE chunks.
1923 		 */
1924 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1925 
1926 		if (maxsize != bp->b_kvasize) {
1927 			vm_offset_t addr = 0;
1928 
1929 			bfreekva(bp);
1930 
1931 			if (vm_map_findspace(buffer_map,
1932 				vm_map_min(buffer_map), maxsize, &addr)) {
1933 				/*
1934 				 * Uh oh.  Buffer map is to fragmented.  We
1935 				 * must defragment the map.
1936 				 */
1937 				atomic_add_int(&bufdefragcnt, 1);
1938 				defrag = 1;
1939 				bp->b_flags |= B_INVAL;
1940 				brelse(bp);
1941 				goto restart;
1942 			}
1943 			if (addr) {
1944 				vm_map_insert(buffer_map, NULL, 0,
1945 					addr, addr + maxsize,
1946 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1947 
1948 				bp->b_kvabase = (caddr_t) addr;
1949 				bp->b_kvasize = maxsize;
1950 				atomic_add_int(&bufspace, bp->b_kvasize);
1951 				atomic_add_int(&bufreusecnt, 1);
1952 			}
1953 		}
1954 		bp->b_saveaddr = bp->b_kvabase;
1955 		bp->b_data = bp->b_saveaddr;
1956 	}
1957 	return(bp);
1958 }
1959 
1960 /*
1961  *	buf_daemon:
1962  *
1963  *	buffer flushing daemon.  Buffers are normally flushed by the
1964  *	update daemon but if it cannot keep up this process starts to
1965  *	take the load in an attempt to prevent getnewbuf() from blocking.
1966  */
1967 
1968 static struct kproc_desc buf_kp = {
1969 	"bufdaemon",
1970 	buf_daemon,
1971 	&bufdaemonproc
1972 };
1973 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1974 
1975 static void
1976 buf_daemon()
1977 {
1978 	int s;
1979 
1980 	mtx_lock(&Giant);
1981 
1982 	/*
1983 	 * This process needs to be suspended prior to shutdown sync.
1984 	 */
1985 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
1986 	    SHUTDOWN_PRI_LAST);
1987 
1988 	/*
1989 	 * This process is allowed to take the buffer cache to the limit
1990 	 */
1991 	s = splbio();
1992 	mtx_lock(&bdlock);
1993 
1994 	for (;;) {
1995 		bd_request = 0;
1996 		mtx_unlock(&bdlock);
1997 
1998 		kthread_suspend_check(bufdaemonproc);
1999 
2000 		/*
2001 		 * Do the flush.  Limit the amount of in-transit I/O we
2002 		 * allow to build up, otherwise we would completely saturate
2003 		 * the I/O system.  Wakeup any waiting processes before we
2004 		 * normally would so they can run in parallel with our drain.
2005 		 */
2006 		while (numdirtybuffers > lodirtybuffers) {
2007 			if (flushbufqueues(0) == 0) {
2008 				/*
2009 				 * Could not find any buffers without rollback
2010 				 * dependencies, so just write the first one
2011 				 * in the hopes of eventually making progress.
2012 				 */
2013 				flushbufqueues(1);
2014 				break;
2015 			}
2016 			waitrunningbufspace();
2017 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2018 		}
2019 
2020 		/*
2021 		 * Only clear bd_request if we have reached our low water
2022 		 * mark.  The buf_daemon normally waits 1 second and
2023 		 * then incrementally flushes any dirty buffers that have
2024 		 * built up, within reason.
2025 		 *
2026 		 * If we were unable to hit our low water mark and couldn't
2027 		 * find any flushable buffers, we sleep half a second.
2028 		 * Otherwise we loop immediately.
2029 		 */
2030 		mtx_lock(&bdlock);
2031 		if (numdirtybuffers <= lodirtybuffers) {
2032 			/*
2033 			 * We reached our low water mark, reset the
2034 			 * request and sleep until we are needed again.
2035 			 * The sleep is just so the suspend code works.
2036 			 */
2037 			bd_request = 0;
2038 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2039 		} else {
2040 			/*
2041 			 * We couldn't find any flushable dirty buffers but
2042 			 * still have too many dirty buffers, we
2043 			 * have to sleep and try again.  (rare)
2044 			 */
2045 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2046 		}
2047 	}
2048 }
2049 
2050 /*
2051  *	flushbufqueues:
2052  *
2053  *	Try to flush a buffer in the dirty queue.  We must be careful to
2054  *	free up B_INVAL buffers instead of write them, which NFS is
2055  *	particularly sensitive to.
2056  */
2057 static int flushwithdeps = 0;
2058 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2059     0, "Number of buffers flushed with dependecies that require rollbacks");
2060 
2061 static int
2062 flushbufqueues(int flushdeps)
2063 {
2064 	struct thread *td = curthread;
2065 	struct vnode *vp;
2066 	struct mount *mp;
2067 	struct buf *bp;
2068 	int hasdeps;
2069 
2070 	mtx_lock(&bqlock);
2071 	TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) {
2072 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2073 			continue;
2074 		BO_LOCK(bp->b_bufobj);
2075 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2076 		    (bp->b_flags & B_DELWRI) == 0) {
2077 			BO_UNLOCK(bp->b_bufobj);
2078 			BUF_UNLOCK(bp);
2079 			continue;
2080 		}
2081 		BO_UNLOCK(bp->b_bufobj);
2082 		if (bp->b_flags & B_INVAL) {
2083 			bremfreel(bp);
2084 			mtx_unlock(&bqlock);
2085 			brelse(bp);
2086 			return (1);
2087 		}
2088 
2089 		if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) {
2090 			if (flushdeps == 0) {
2091 				BUF_UNLOCK(bp);
2092 				continue;
2093 			}
2094 			hasdeps = 1;
2095 		} else
2096 			hasdeps = 0;
2097 		/*
2098 		 * We must hold the lock on a vnode before writing
2099 		 * one of its buffers. Otherwise we may confuse, or
2100 		 * in the case of a snapshot vnode, deadlock the
2101 		 * system.
2102 		 *
2103 		 * The lock order here is the reverse of the normal
2104 		 * of vnode followed by buf lock.  This is ok because
2105 		 * the NOWAIT will prevent deadlock.
2106 		 */
2107 		vp = bp->b_vp;
2108 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2109 			BUF_UNLOCK(bp);
2110 			continue;
2111 		}
2112 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2113 			mtx_unlock(&bqlock);
2114 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2115 			    bp, bp->b_vp, bp->b_flags);
2116 			vfs_bio_awrite(bp);
2117 			vn_finished_write(mp);
2118 			VOP_UNLOCK(vp, 0, td);
2119 			flushwithdeps += hasdeps;
2120 			return (1);
2121 		}
2122 		vn_finished_write(mp);
2123 		BUF_UNLOCK(bp);
2124 	}
2125 	mtx_unlock(&bqlock);
2126 	return (0);
2127 }
2128 
2129 /*
2130  * Check to see if a block is currently memory resident.
2131  */
2132 struct buf *
2133 incore(struct bufobj *bo, daddr_t blkno)
2134 {
2135 	struct buf *bp;
2136 
2137 	int s = splbio();
2138 	BO_LOCK(bo);
2139 	bp = gbincore(bo, blkno);
2140 	BO_UNLOCK(bo);
2141 	splx(s);
2142 	return (bp);
2143 }
2144 
2145 /*
2146  * Returns true if no I/O is needed to access the
2147  * associated VM object.  This is like incore except
2148  * it also hunts around in the VM system for the data.
2149  */
2150 
2151 static int
2152 inmem(struct vnode * vp, daddr_t blkno)
2153 {
2154 	vm_object_t obj;
2155 	vm_offset_t toff, tinc, size;
2156 	vm_page_t m;
2157 	vm_ooffset_t off;
2158 
2159 	ASSERT_VOP_LOCKED(vp, "inmem");
2160 
2161 	if (incore(&vp->v_bufobj, blkno))
2162 		return 1;
2163 	if (vp->v_mount == NULL)
2164 		return 0;
2165 	obj = vp->v_object;
2166 	if (obj == NULL)
2167 		return (0);
2168 
2169 	size = PAGE_SIZE;
2170 	if (size > vp->v_mount->mnt_stat.f_iosize)
2171 		size = vp->v_mount->mnt_stat.f_iosize;
2172 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2173 
2174 	VM_OBJECT_LOCK(obj);
2175 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2176 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2177 		if (!m)
2178 			goto notinmem;
2179 		tinc = size;
2180 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2181 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2182 		if (vm_page_is_valid(m,
2183 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2184 			goto notinmem;
2185 	}
2186 	VM_OBJECT_UNLOCK(obj);
2187 	return 1;
2188 
2189 notinmem:
2190 	VM_OBJECT_UNLOCK(obj);
2191 	return (0);
2192 }
2193 
2194 /*
2195  *	vfs_setdirty:
2196  *
2197  *	Sets the dirty range for a buffer based on the status of the dirty
2198  *	bits in the pages comprising the buffer.
2199  *
2200  *	The range is limited to the size of the buffer.
2201  *
2202  *	This routine is primarily used by NFS, but is generalized for the
2203  *	B_VMIO case.
2204  */
2205 static void
2206 vfs_setdirty(struct buf *bp)
2207 {
2208 	int i;
2209 	vm_object_t object;
2210 
2211 	/*
2212 	 * Degenerate case - empty buffer
2213 	 */
2214 
2215 	if (bp->b_bufsize == 0)
2216 		return;
2217 
2218 	/*
2219 	 * We qualify the scan for modified pages on whether the
2220 	 * object has been flushed yet.  The OBJ_WRITEABLE flag
2221 	 * is not cleared simply by protecting pages off.
2222 	 */
2223 
2224 	if ((bp->b_flags & B_VMIO) == 0)
2225 		return;
2226 
2227 	object = bp->b_pages[0]->object;
2228 	VM_OBJECT_LOCK(object);
2229 	if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2230 		printf("Warning: object %p writeable but not mightbedirty\n", object);
2231 	if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2232 		printf("Warning: object %p mightbedirty but not writeable\n", object);
2233 
2234 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2235 		vm_offset_t boffset;
2236 		vm_offset_t eoffset;
2237 
2238 		vm_page_lock_queues();
2239 		/*
2240 		 * test the pages to see if they have been modified directly
2241 		 * by users through the VM system.
2242 		 */
2243 		for (i = 0; i < bp->b_npages; i++)
2244 			vm_page_test_dirty(bp->b_pages[i]);
2245 
2246 		/*
2247 		 * Calculate the encompassing dirty range, boffset and eoffset,
2248 		 * (eoffset - boffset) bytes.
2249 		 */
2250 
2251 		for (i = 0; i < bp->b_npages; i++) {
2252 			if (bp->b_pages[i]->dirty)
2253 				break;
2254 		}
2255 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2256 
2257 		for (i = bp->b_npages - 1; i >= 0; --i) {
2258 			if (bp->b_pages[i]->dirty) {
2259 				break;
2260 			}
2261 		}
2262 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2263 
2264 		vm_page_unlock_queues();
2265 		/*
2266 		 * Fit it to the buffer.
2267 		 */
2268 
2269 		if (eoffset > bp->b_bcount)
2270 			eoffset = bp->b_bcount;
2271 
2272 		/*
2273 		 * If we have a good dirty range, merge with the existing
2274 		 * dirty range.
2275 		 */
2276 
2277 		if (boffset < eoffset) {
2278 			if (bp->b_dirtyoff > boffset)
2279 				bp->b_dirtyoff = boffset;
2280 			if (bp->b_dirtyend < eoffset)
2281 				bp->b_dirtyend = eoffset;
2282 		}
2283 	}
2284 	VM_OBJECT_UNLOCK(object);
2285 }
2286 
2287 /*
2288  *	getblk:
2289  *
2290  *	Get a block given a specified block and offset into a file/device.
2291  *	The buffers B_DONE bit will be cleared on return, making it almost
2292  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2293  *	return.  The caller should clear B_INVAL prior to initiating a
2294  *	READ.
2295  *
2296  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2297  *	an existing buffer.
2298  *
2299  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2300  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2301  *	and then cleared based on the backing VM.  If the previous buffer is
2302  *	non-0-sized but invalid, B_CACHE will be cleared.
2303  *
2304  *	If getblk() must create a new buffer, the new buffer is returned with
2305  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2306  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2307  *	backing VM.
2308  *
2309  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2310  *	B_CACHE bit is clear.
2311  *
2312  *	What this means, basically, is that the caller should use B_CACHE to
2313  *	determine whether the buffer is fully valid or not and should clear
2314  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2315  *	the buffer by loading its data area with something, the caller needs
2316  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2317  *	the caller should set B_CACHE ( as an optimization ), else the caller
2318  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2319  *	a write attempt or if it was a successfull read.  If the caller
2320  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2321  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2322  */
2323 struct buf *
2324 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2325     int flags)
2326 {
2327 	struct buf *bp;
2328 	struct bufobj *bo;
2329 	int s;
2330 	int error;
2331 
2332 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2333 	ASSERT_VOP_LOCKED(vp, "getblk");
2334 	if (size > MAXBSIZE)
2335 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2336 
2337 	bo = &vp->v_bufobj;
2338 	s = splbio();
2339 loop:
2340 	/*
2341 	 * Block if we are low on buffers.   Certain processes are allowed
2342 	 * to completely exhaust the buffer cache.
2343          *
2344          * If this check ever becomes a bottleneck it may be better to
2345          * move it into the else, when gbincore() fails.  At the moment
2346          * it isn't a problem.
2347 	 *
2348 	 * XXX remove if 0 sections (clean this up after its proven)
2349          */
2350 	if (numfreebuffers == 0) {
2351 		if (curthread == PCPU_GET(idlethread))
2352 			return NULL;
2353 		mtx_lock(&nblock);
2354 		needsbuffer |= VFS_BIO_NEED_ANY;
2355 		mtx_unlock(&nblock);
2356 	}
2357 
2358 	VI_LOCK(vp);
2359 	bp = gbincore(bo, blkno);
2360 	if (bp != NULL) {
2361 		int lockflags;
2362 		/*
2363 		 * Buffer is in-core.  If the buffer is not busy, it must
2364 		 * be on a queue.
2365 		 */
2366 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2367 
2368 		if (flags & GB_LOCK_NOWAIT)
2369 			lockflags |= LK_NOWAIT;
2370 
2371 		error = BUF_TIMELOCK(bp, lockflags,
2372 		    VI_MTX(vp), "getblk", slpflag, slptimeo);
2373 
2374 		/*
2375 		 * If we slept and got the lock we have to restart in case
2376 		 * the buffer changed identities.
2377 		 */
2378 		if (error == ENOLCK)
2379 			goto loop;
2380 		/* We timed out or were interrupted. */
2381 		else if (error)
2382 			return (NULL);
2383 
2384 		/*
2385 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2386 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2387 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2388 		 * backing VM cache.
2389 		 */
2390 		if (bp->b_flags & B_INVAL)
2391 			bp->b_flags &= ~B_CACHE;
2392 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2393 			bp->b_flags |= B_CACHE;
2394 		bremfree(bp);
2395 
2396 		/*
2397 		 * check for size inconsistancies for non-VMIO case.
2398 		 */
2399 
2400 		if (bp->b_bcount != size) {
2401 			if ((bp->b_flags & B_VMIO) == 0 ||
2402 			    (size > bp->b_kvasize)) {
2403 				if (bp->b_flags & B_DELWRI) {
2404 					bp->b_flags |= B_NOCACHE;
2405 					bwrite(bp);
2406 				} else {
2407 					if ((bp->b_flags & B_VMIO) &&
2408 					   (LIST_FIRST(&bp->b_dep) == NULL)) {
2409 						bp->b_flags |= B_RELBUF;
2410 						brelse(bp);
2411 					} else {
2412 						bp->b_flags |= B_NOCACHE;
2413 						bwrite(bp);
2414 					}
2415 				}
2416 				goto loop;
2417 			}
2418 		}
2419 
2420 		/*
2421 		 * If the size is inconsistant in the VMIO case, we can resize
2422 		 * the buffer.  This might lead to B_CACHE getting set or
2423 		 * cleared.  If the size has not changed, B_CACHE remains
2424 		 * unchanged from its previous state.
2425 		 */
2426 
2427 		if (bp->b_bcount != size)
2428 			allocbuf(bp, size);
2429 
2430 		KASSERT(bp->b_offset != NOOFFSET,
2431 		    ("getblk: no buffer offset"));
2432 
2433 		/*
2434 		 * A buffer with B_DELWRI set and B_CACHE clear must
2435 		 * be committed before we can return the buffer in
2436 		 * order to prevent the caller from issuing a read
2437 		 * ( due to B_CACHE not being set ) and overwriting
2438 		 * it.
2439 		 *
2440 		 * Most callers, including NFS and FFS, need this to
2441 		 * operate properly either because they assume they
2442 		 * can issue a read if B_CACHE is not set, or because
2443 		 * ( for example ) an uncached B_DELWRI might loop due
2444 		 * to softupdates re-dirtying the buffer.  In the latter
2445 		 * case, B_CACHE is set after the first write completes,
2446 		 * preventing further loops.
2447 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2448 		 * above while extending the buffer, we cannot allow the
2449 		 * buffer to remain with B_CACHE set after the write
2450 		 * completes or it will represent a corrupt state.  To
2451 		 * deal with this we set B_NOCACHE to scrap the buffer
2452 		 * after the write.
2453 		 *
2454 		 * We might be able to do something fancy, like setting
2455 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2456 		 * so the below call doesn't set B_CACHE, but that gets real
2457 		 * confusing.  This is much easier.
2458 		 */
2459 
2460 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2461 			bp->b_flags |= B_NOCACHE;
2462 			bwrite(bp);
2463 			goto loop;
2464 		}
2465 
2466 		splx(s);
2467 		bp->b_flags &= ~B_DONE;
2468 	} else {
2469 		int bsize, maxsize, vmio;
2470 		off_t offset;
2471 
2472 		/*
2473 		 * Buffer is not in-core, create new buffer.  The buffer
2474 		 * returned by getnewbuf() is locked.  Note that the returned
2475 		 * buffer is also considered valid (not marked B_INVAL).
2476 		 */
2477 		VI_UNLOCK(vp);
2478 		/*
2479 		 * If the user does not want us to create the buffer, bail out
2480 		 * here.
2481 		 */
2482 		if (flags & GB_NOCREAT) {
2483 			splx(s);
2484 			return NULL;
2485 		}
2486 
2487 		bsize = bo->bo_bsize;
2488 		offset = blkno * bsize;
2489 		vmio = vp->v_object != NULL;
2490 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2491 		maxsize = imax(maxsize, bsize);
2492 
2493 		bp = getnewbuf(slpflag, slptimeo, size, maxsize);
2494 		if (bp == NULL) {
2495 			if (slpflag || slptimeo) {
2496 				splx(s);
2497 				return NULL;
2498 			}
2499 			goto loop;
2500 		}
2501 
2502 		/*
2503 		 * This code is used to make sure that a buffer is not
2504 		 * created while the getnewbuf routine is blocked.
2505 		 * This can be a problem whether the vnode is locked or not.
2506 		 * If the buffer is created out from under us, we have to
2507 		 * throw away the one we just created.  There is now window
2508 		 * race because we are safely running at splbio() from the
2509 		 * point of the duplicate buffer creation through to here,
2510 		 * and we've locked the buffer.
2511 		 *
2512 		 * Note: this must occur before we associate the buffer
2513 		 * with the vp especially considering limitations in
2514 		 * the splay tree implementation when dealing with duplicate
2515 		 * lblkno's.
2516 		 */
2517 		BO_LOCK(bo);
2518 		if (gbincore(bo, blkno)) {
2519 			BO_UNLOCK(bo);
2520 			bp->b_flags |= B_INVAL;
2521 			brelse(bp);
2522 			goto loop;
2523 		}
2524 
2525 		/*
2526 		 * Insert the buffer into the hash, so that it can
2527 		 * be found by incore.
2528 		 */
2529 		bp->b_blkno = bp->b_lblkno = blkno;
2530 		bp->b_offset = offset;
2531 
2532 		bgetvp(vp, bp);
2533 		BO_UNLOCK(bo);
2534 
2535 		/*
2536 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2537 		 * buffer size starts out as 0, B_CACHE will be set by
2538 		 * allocbuf() for the VMIO case prior to it testing the
2539 		 * backing store for validity.
2540 		 */
2541 
2542 		if (vmio) {
2543 			bp->b_flags |= B_VMIO;
2544 #if defined(VFS_BIO_DEBUG)
2545 			if (vn_canvmio(vp) != TRUE)
2546 				printf("getblk: VMIO on vnode type %d\n",
2547 					vp->v_type);
2548 #endif
2549 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2550 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2551 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2552 		} else {
2553 			bp->b_flags &= ~B_VMIO;
2554 			KASSERT(bp->b_bufobj->bo_object == NULL,
2555 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2556 			    bp, bp->b_bufobj->bo_object));
2557 		}
2558 
2559 		allocbuf(bp, size);
2560 
2561 		splx(s);
2562 		bp->b_flags &= ~B_DONE;
2563 	}
2564 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2565 	KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2566 	KASSERT(bp->b_bufobj == bo,
2567 	    ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo));
2568 	return (bp);
2569 }
2570 
2571 /*
2572  * Get an empty, disassociated buffer of given size.  The buffer is initially
2573  * set to B_INVAL.
2574  */
2575 struct buf *
2576 geteblk(int size)
2577 {
2578 	struct buf *bp;
2579 	int s;
2580 	int maxsize;
2581 
2582 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2583 
2584 	s = splbio();
2585 	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2586 		continue;
2587 	splx(s);
2588 	allocbuf(bp, size);
2589 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2590 	KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2591 	return (bp);
2592 }
2593 
2594 
2595 /*
2596  * This code constitutes the buffer memory from either anonymous system
2597  * memory (in the case of non-VMIO operations) or from an associated
2598  * VM object (in the case of VMIO operations).  This code is able to
2599  * resize a buffer up or down.
2600  *
2601  * Note that this code is tricky, and has many complications to resolve
2602  * deadlock or inconsistant data situations.  Tread lightly!!!
2603  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2604  * the caller.  Calling this code willy nilly can result in the loss of data.
2605  *
2606  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2607  * B_CACHE for the non-VMIO case.
2608  */
2609 
2610 int
2611 allocbuf(struct buf *bp, int size)
2612 {
2613 	int newbsize, mbsize;
2614 	int i;
2615 
2616 	if (BUF_REFCNT(bp) == 0)
2617 		panic("allocbuf: buffer not busy");
2618 
2619 	if (bp->b_kvasize < size)
2620 		panic("allocbuf: buffer too small");
2621 
2622 	if ((bp->b_flags & B_VMIO) == 0) {
2623 		caddr_t origbuf;
2624 		int origbufsize;
2625 		/*
2626 		 * Just get anonymous memory from the kernel.  Don't
2627 		 * mess with B_CACHE.
2628 		 */
2629 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2630 		if (bp->b_flags & B_MALLOC)
2631 			newbsize = mbsize;
2632 		else
2633 			newbsize = round_page(size);
2634 
2635 		if (newbsize < bp->b_bufsize) {
2636 			/*
2637 			 * malloced buffers are not shrunk
2638 			 */
2639 			if (bp->b_flags & B_MALLOC) {
2640 				if (newbsize) {
2641 					bp->b_bcount = size;
2642 				} else {
2643 					free(bp->b_data, M_BIOBUF);
2644 					if (bp->b_bufsize) {
2645 						atomic_subtract_int(
2646 						    &bufmallocspace,
2647 						    bp->b_bufsize);
2648 						bufspacewakeup();
2649 						bp->b_bufsize = 0;
2650 					}
2651 					bp->b_saveaddr = bp->b_kvabase;
2652 					bp->b_data = bp->b_saveaddr;
2653 					bp->b_bcount = 0;
2654 					bp->b_flags &= ~B_MALLOC;
2655 				}
2656 				return 1;
2657 			}
2658 			vm_hold_free_pages(
2659 			    bp,
2660 			    (vm_offset_t) bp->b_data + newbsize,
2661 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2662 		} else if (newbsize > bp->b_bufsize) {
2663 			/*
2664 			 * We only use malloced memory on the first allocation.
2665 			 * and revert to page-allocated memory when the buffer
2666 			 * grows.
2667 			 */
2668 			/*
2669 			 * There is a potential smp race here that could lead
2670 			 * to bufmallocspace slightly passing the max.  It
2671 			 * is probably extremely rare and not worth worrying
2672 			 * over.
2673 			 */
2674 			if ( (bufmallocspace < maxbufmallocspace) &&
2675 				(bp->b_bufsize == 0) &&
2676 				(mbsize <= PAGE_SIZE/2)) {
2677 
2678 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2679 				bp->b_bufsize = mbsize;
2680 				bp->b_bcount = size;
2681 				bp->b_flags |= B_MALLOC;
2682 				atomic_add_int(&bufmallocspace, mbsize);
2683 				return 1;
2684 			}
2685 			origbuf = NULL;
2686 			origbufsize = 0;
2687 			/*
2688 			 * If the buffer is growing on its other-than-first allocation,
2689 			 * then we revert to the page-allocation scheme.
2690 			 */
2691 			if (bp->b_flags & B_MALLOC) {
2692 				origbuf = bp->b_data;
2693 				origbufsize = bp->b_bufsize;
2694 				bp->b_data = bp->b_kvabase;
2695 				if (bp->b_bufsize) {
2696 					atomic_subtract_int(&bufmallocspace,
2697 					    bp->b_bufsize);
2698 					bufspacewakeup();
2699 					bp->b_bufsize = 0;
2700 				}
2701 				bp->b_flags &= ~B_MALLOC;
2702 				newbsize = round_page(newbsize);
2703 			}
2704 			vm_hold_load_pages(
2705 			    bp,
2706 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2707 			    (vm_offset_t) bp->b_data + newbsize);
2708 			if (origbuf) {
2709 				bcopy(origbuf, bp->b_data, origbufsize);
2710 				free(origbuf, M_BIOBUF);
2711 			}
2712 		}
2713 	} else {
2714 		int desiredpages;
2715 
2716 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2717 		desiredpages = (size == 0) ? 0 :
2718 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2719 
2720 		if (bp->b_flags & B_MALLOC)
2721 			panic("allocbuf: VMIO buffer can't be malloced");
2722 		/*
2723 		 * Set B_CACHE initially if buffer is 0 length or will become
2724 		 * 0-length.
2725 		 */
2726 		if (size == 0 || bp->b_bufsize == 0)
2727 			bp->b_flags |= B_CACHE;
2728 
2729 		if (newbsize < bp->b_bufsize) {
2730 			/*
2731 			 * DEV_BSIZE aligned new buffer size is less then the
2732 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2733 			 * if we have to remove any pages.
2734 			 */
2735 			if (desiredpages < bp->b_npages) {
2736 				vm_page_t m;
2737 
2738 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2739 				vm_page_lock_queues();
2740 				for (i = desiredpages; i < bp->b_npages; i++) {
2741 					/*
2742 					 * the page is not freed here -- it
2743 					 * is the responsibility of
2744 					 * vnode_pager_setsize
2745 					 */
2746 					m = bp->b_pages[i];
2747 					KASSERT(m != bogus_page,
2748 					    ("allocbuf: bogus page found"));
2749 					while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2750 						vm_page_lock_queues();
2751 
2752 					bp->b_pages[i] = NULL;
2753 					vm_page_unwire(m, 0);
2754 				}
2755 				vm_page_unlock_queues();
2756 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2757 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2758 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2759 				bp->b_npages = desiredpages;
2760 			}
2761 		} else if (size > bp->b_bcount) {
2762 			/*
2763 			 * We are growing the buffer, possibly in a
2764 			 * byte-granular fashion.
2765 			 */
2766 			struct vnode *vp;
2767 			vm_object_t obj;
2768 			vm_offset_t toff;
2769 			vm_offset_t tinc;
2770 
2771 			/*
2772 			 * Step 1, bring in the VM pages from the object,
2773 			 * allocating them if necessary.  We must clear
2774 			 * B_CACHE if these pages are not valid for the
2775 			 * range covered by the buffer.
2776 			 */
2777 
2778 			vp = bp->b_vp;
2779 			obj = bp->b_bufobj->bo_object;
2780 
2781 			VM_OBJECT_LOCK(obj);
2782 			while (bp->b_npages < desiredpages) {
2783 				vm_page_t m;
2784 				vm_pindex_t pi;
2785 
2786 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2787 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2788 					/*
2789 					 * note: must allocate system pages
2790 					 * since blocking here could intefere
2791 					 * with paging I/O, no matter which
2792 					 * process we are.
2793 					 */
2794 					m = vm_page_alloc(obj, pi,
2795 					    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2796 					    VM_ALLOC_WIRED);
2797 					if (m == NULL) {
2798 						atomic_add_int(&vm_pageout_deficit,
2799 						    desiredpages - bp->b_npages);
2800 						VM_OBJECT_UNLOCK(obj);
2801 						VM_WAIT;
2802 						VM_OBJECT_LOCK(obj);
2803 					} else {
2804 						bp->b_flags &= ~B_CACHE;
2805 						bp->b_pages[bp->b_npages] = m;
2806 						++bp->b_npages;
2807 					}
2808 					continue;
2809 				}
2810 
2811 				/*
2812 				 * We found a page.  If we have to sleep on it,
2813 				 * retry because it might have gotten freed out
2814 				 * from under us.
2815 				 *
2816 				 * We can only test PG_BUSY here.  Blocking on
2817 				 * m->busy might lead to a deadlock:
2818 				 *
2819 				 *  vm_fault->getpages->cluster_read->allocbuf
2820 				 *
2821 				 */
2822 				vm_page_lock_queues();
2823 				if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2824 					continue;
2825 
2826 				/*
2827 				 * We have a good page.  Should we wakeup the
2828 				 * page daemon?
2829 				 */
2830 				if ((curproc != pageproc) &&
2831 				    ((m->queue - m->pc) == PQ_CACHE) &&
2832 				    ((cnt.v_free_count + cnt.v_cache_count) <
2833 					(cnt.v_free_min + cnt.v_cache_min))) {
2834 					pagedaemon_wakeup();
2835 				}
2836 				vm_page_wire(m);
2837 				vm_page_unlock_queues();
2838 				bp->b_pages[bp->b_npages] = m;
2839 				++bp->b_npages;
2840 			}
2841 
2842 			/*
2843 			 * Step 2.  We've loaded the pages into the buffer,
2844 			 * we have to figure out if we can still have B_CACHE
2845 			 * set.  Note that B_CACHE is set according to the
2846 			 * byte-granular range ( bcount and size ), new the
2847 			 * aligned range ( newbsize ).
2848 			 *
2849 			 * The VM test is against m->valid, which is DEV_BSIZE
2850 			 * aligned.  Needless to say, the validity of the data
2851 			 * needs to also be DEV_BSIZE aligned.  Note that this
2852 			 * fails with NFS if the server or some other client
2853 			 * extends the file's EOF.  If our buffer is resized,
2854 			 * B_CACHE may remain set! XXX
2855 			 */
2856 
2857 			toff = bp->b_bcount;
2858 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2859 
2860 			while ((bp->b_flags & B_CACHE) && toff < size) {
2861 				vm_pindex_t pi;
2862 
2863 				if (tinc > (size - toff))
2864 					tinc = size - toff;
2865 
2866 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2867 				    PAGE_SHIFT;
2868 
2869 				vfs_buf_test_cache(
2870 				    bp,
2871 				    bp->b_offset,
2872 				    toff,
2873 				    tinc,
2874 				    bp->b_pages[pi]
2875 				);
2876 				toff += tinc;
2877 				tinc = PAGE_SIZE;
2878 			}
2879 			VM_OBJECT_UNLOCK(obj);
2880 
2881 			/*
2882 			 * Step 3, fixup the KVM pmap.  Remember that
2883 			 * bp->b_data is relative to bp->b_offset, but
2884 			 * bp->b_offset may be offset into the first page.
2885 			 */
2886 
2887 			bp->b_data = (caddr_t)
2888 			    trunc_page((vm_offset_t)bp->b_data);
2889 			pmap_qenter(
2890 			    (vm_offset_t)bp->b_data,
2891 			    bp->b_pages,
2892 			    bp->b_npages
2893 			);
2894 
2895 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2896 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
2897 		}
2898 	}
2899 	if (newbsize < bp->b_bufsize)
2900 		bufspacewakeup();
2901 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
2902 	bp->b_bcount = size;		/* requested buffer size	*/
2903 	return 1;
2904 }
2905 
2906 void
2907 biodone(struct bio *bp)
2908 {
2909 
2910 	mtx_lock(&bdonelock);
2911 	bp->bio_flags |= BIO_DONE;
2912 	if (bp->bio_done == NULL)
2913 		wakeup(bp);
2914 	mtx_unlock(&bdonelock);
2915 	if (bp->bio_done != NULL)
2916 		bp->bio_done(bp);
2917 }
2918 
2919 /*
2920  * Wait for a BIO to finish.
2921  *
2922  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
2923  * case is not yet clear.
2924  */
2925 int
2926 biowait(struct bio *bp, const char *wchan)
2927 {
2928 
2929 	mtx_lock(&bdonelock);
2930 	while ((bp->bio_flags & BIO_DONE) == 0)
2931 		msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
2932 	mtx_unlock(&bdonelock);
2933 	if (bp->bio_error != 0)
2934 		return (bp->bio_error);
2935 	if (!(bp->bio_flags & BIO_ERROR))
2936 		return (0);
2937 	return (EIO);
2938 }
2939 
2940 void
2941 biofinish(struct bio *bp, struct devstat *stat, int error)
2942 {
2943 
2944 	if (error) {
2945 		bp->bio_error = error;
2946 		bp->bio_flags |= BIO_ERROR;
2947 	}
2948 	if (stat != NULL)
2949 		devstat_end_transaction_bio(stat, bp);
2950 	biodone(bp);
2951 }
2952 
2953 /*
2954  *	bufwait:
2955  *
2956  *	Wait for buffer I/O completion, returning error status.  The buffer
2957  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
2958  *	error and cleared.
2959  */
2960 int
2961 bufwait(struct buf *bp)
2962 {
2963 	int s;
2964 
2965 	s = splbio();
2966 	if (bp->b_iocmd == BIO_READ)
2967 		bwait(bp, PRIBIO, "biord");
2968 	else
2969 		bwait(bp, PRIBIO, "biowr");
2970 	splx(s);
2971 	if (bp->b_flags & B_EINTR) {
2972 		bp->b_flags &= ~B_EINTR;
2973 		return (EINTR);
2974 	}
2975 	if (bp->b_ioflags & BIO_ERROR) {
2976 		return (bp->b_error ? bp->b_error : EIO);
2977 	} else {
2978 		return (0);
2979 	}
2980 }
2981 
2982  /*
2983   * Call back function from struct bio back up to struct buf.
2984   */
2985 static void
2986 bufdonebio(struct bio *bip)
2987 {
2988 	struct buf *bp;
2989 
2990 	bp = bip->bio_caller2;
2991 	bp->b_resid = bp->b_bcount - bip->bio_completed;
2992 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
2993 	bp->b_ioflags = bip->bio_flags;
2994 	bp->b_error = bip->bio_error;
2995 	if (bp->b_error)
2996 		bp->b_ioflags |= BIO_ERROR;
2997 	bufdone(bp);
2998 	g_destroy_bio(bip);
2999 }
3000 
3001 void
3002 dev_strategy(struct cdev *dev, struct buf *bp)
3003 {
3004 	struct cdevsw *csw;
3005 	struct bio *bip;
3006 
3007 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3008 		panic("b_iocmd botch");
3009 	for (;;) {
3010 		bip = g_new_bio();
3011 		if (bip != NULL)
3012 			break;
3013 		/* Try again later */
3014 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3015 	}
3016 	bip->bio_cmd = bp->b_iocmd;
3017 	bip->bio_offset = bp->b_iooffset;
3018 	bip->bio_length = bp->b_bcount;
3019 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3020 	bip->bio_data = bp->b_data;
3021 	bip->bio_done = bufdonebio;
3022 	bip->bio_caller2 = bp;
3023 	bip->bio_dev = dev;
3024 	KASSERT(dev->si_refcount > 0,
3025 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3026 	    devtoname(dev)));
3027 	csw = dev_refthread(dev);
3028 	if (csw == NULL) {
3029 		bp->b_error = ENXIO;
3030 		bp->b_ioflags = BIO_ERROR;
3031 		bufdone(bp);
3032 		return;
3033 	}
3034 	(*csw->d_strategy)(bip);
3035 	dev_relthread(dev);
3036 }
3037 
3038 /*
3039  *	bufdone:
3040  *
3041  *	Finish I/O on a buffer, optionally calling a completion function.
3042  *	This is usually called from an interrupt so process blocking is
3043  *	not allowed.
3044  *
3045  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3046  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3047  *	assuming B_INVAL is clear.
3048  *
3049  *	For the VMIO case, we set B_CACHE if the op was a read and no
3050  *	read error occured, or if the op was a write.  B_CACHE is never
3051  *	set if the buffer is invalid or otherwise uncacheable.
3052  *
3053  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3054  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3055  *	in the biodone routine.
3056  */
3057 void
3058 bufdone(struct buf *bp)
3059 {
3060 	struct bufobj *dropobj;
3061 	int s;
3062 	void    (*biodone)(struct buf *);
3063 
3064 
3065 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3066 	s = splbio();
3067 	dropobj = NULL;
3068 
3069 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
3070 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3071 
3072 	runningbufwakeup(bp);
3073 	if (bp->b_iocmd == BIO_WRITE)
3074 		dropobj = bp->b_bufobj;
3075 	/* call optional completion function if requested */
3076 	if (bp->b_iodone != NULL) {
3077 		biodone = bp->b_iodone;
3078 		bp->b_iodone = NULL;
3079 		/*
3080 		 * Device drivers may or may not hold giant, hold it here
3081 		 * if we're calling into unknown code.
3082 		 */
3083 		mtx_lock(&Giant);
3084 		bp->b_flags |= B_DONE;
3085 		(*biodone) (bp);
3086 		mtx_unlock(&Giant);
3087 		if (dropobj)
3088 			bufobj_wdrop(dropobj);
3089 		splx(s);
3090 		return;
3091 	}
3092 	if (LIST_FIRST(&bp->b_dep) != NULL)
3093 		buf_complete(bp);
3094 
3095 	if (bp->b_flags & B_VMIO) {
3096 		int i;
3097 		vm_ooffset_t foff;
3098 		vm_page_t m;
3099 		vm_object_t obj;
3100 		int iosize;
3101 		struct vnode *vp = bp->b_vp;
3102 
3103 		obj = bp->b_bufobj->bo_object;
3104 
3105 #if defined(VFS_BIO_DEBUG)
3106 		mp_fixme("usecount and vflag accessed without locks.");
3107 		if (vp->v_usecount == 0) {
3108 			panic("biodone: zero vnode ref count");
3109 		}
3110 
3111 		KASSERT(vp->v_object != NULL,
3112 			("biodone: vnode %p has no vm_object", vp));
3113 #endif
3114 
3115 		foff = bp->b_offset;
3116 		KASSERT(bp->b_offset != NOOFFSET,
3117 		    ("biodone: no buffer offset"));
3118 
3119 		VM_OBJECT_LOCK(obj);
3120 #if defined(VFS_BIO_DEBUG)
3121 		if (obj->paging_in_progress < bp->b_npages) {
3122 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3123 			    obj->paging_in_progress, bp->b_npages);
3124 		}
3125 #endif
3126 
3127 		/*
3128 		 * Set B_CACHE if the op was a normal read and no error
3129 		 * occured.  B_CACHE is set for writes in the b*write()
3130 		 * routines.
3131 		 */
3132 		iosize = bp->b_bcount - bp->b_resid;
3133 		if (bp->b_iocmd == BIO_READ &&
3134 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3135 		    !(bp->b_ioflags & BIO_ERROR)) {
3136 			bp->b_flags |= B_CACHE;
3137 		}
3138 		vm_page_lock_queues();
3139 		for (i = 0; i < bp->b_npages; i++) {
3140 			int bogusflag = 0;
3141 			int resid;
3142 
3143 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3144 			if (resid > iosize)
3145 				resid = iosize;
3146 
3147 			/*
3148 			 * cleanup bogus pages, restoring the originals
3149 			 */
3150 			m = bp->b_pages[i];
3151 			if (m == bogus_page) {
3152 				bogusflag = 1;
3153 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3154 				if (m == NULL)
3155 					panic("biodone: page disappeared!");
3156 				bp->b_pages[i] = m;
3157 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3158 			}
3159 #if defined(VFS_BIO_DEBUG)
3160 			if (OFF_TO_IDX(foff) != m->pindex) {
3161 				printf(
3162 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3163 				    (intmax_t)foff, (uintmax_t)m->pindex);
3164 			}
3165 #endif
3166 
3167 			/*
3168 			 * In the write case, the valid and clean bits are
3169 			 * already changed correctly ( see bdwrite() ), so we
3170 			 * only need to do this here in the read case.
3171 			 */
3172 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3173 				vfs_page_set_valid(bp, foff, i, m);
3174 			}
3175 
3176 			/*
3177 			 * when debugging new filesystems or buffer I/O methods, this
3178 			 * is the most common error that pops up.  if you see this, you
3179 			 * have not set the page busy flag correctly!!!
3180 			 */
3181 			if (m->busy == 0) {
3182 				printf("biodone: page busy < 0, "
3183 				    "pindex: %d, foff: 0x(%x,%x), "
3184 				    "resid: %d, index: %d\n",
3185 				    (int) m->pindex, (int)(foff >> 32),
3186 						(int) foff & 0xffffffff, resid, i);
3187 				if (!vn_isdisk(vp, NULL))
3188 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3189 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3190 					    (intmax_t) bp->b_lblkno,
3191 					    bp->b_flags, bp->b_npages);
3192 				else
3193 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3194 					    (intmax_t) bp->b_lblkno,
3195 					    bp->b_flags, bp->b_npages);
3196 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3197 				    (u_long)m->valid, (u_long)m->dirty,
3198 				    m->wire_count);
3199 				panic("biodone: page busy < 0\n");
3200 			}
3201 			vm_page_io_finish(m);
3202 			vm_object_pip_subtract(obj, 1);
3203 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3204 			iosize -= resid;
3205 		}
3206 		vm_page_unlock_queues();
3207 		vm_object_pip_wakeupn(obj, 0);
3208 		VM_OBJECT_UNLOCK(obj);
3209 	}
3210 
3211 	/*
3212 	 * For asynchronous completions, release the buffer now. The brelse
3213 	 * will do a wakeup there if necessary - so no need to do a wakeup
3214 	 * here in the async case. The sync case always needs to do a wakeup.
3215 	 */
3216 
3217 	if (bp->b_flags & B_ASYNC) {
3218 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3219 			brelse(bp);
3220 		else
3221 			bqrelse(bp);
3222 	} else
3223 		bdone(bp);
3224 	if (dropobj)
3225 		bufobj_wdrop(dropobj);
3226 	splx(s);
3227 }
3228 
3229 /*
3230  * This routine is called in lieu of iodone in the case of
3231  * incomplete I/O.  This keeps the busy status for pages
3232  * consistant.
3233  */
3234 void
3235 vfs_unbusy_pages(struct buf *bp)
3236 {
3237 	int i;
3238 	vm_object_t obj;
3239 	vm_page_t m;
3240 
3241 	runningbufwakeup(bp);
3242 	if (!(bp->b_flags & B_VMIO))
3243 		return;
3244 
3245 	obj = bp->b_bufobj->bo_object;
3246 	VM_OBJECT_LOCK(obj);
3247 	vm_page_lock_queues();
3248 	for (i = 0; i < bp->b_npages; i++) {
3249 		m = bp->b_pages[i];
3250 		if (m == bogus_page) {
3251 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3252 			if (!m)
3253 				panic("vfs_unbusy_pages: page missing\n");
3254 			bp->b_pages[i] = m;
3255 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3256 			    bp->b_pages, bp->b_npages);
3257 		}
3258 		vm_object_pip_subtract(obj, 1);
3259 		vm_page_io_finish(m);
3260 	}
3261 	vm_page_unlock_queues();
3262 	vm_object_pip_wakeupn(obj, 0);
3263 	VM_OBJECT_UNLOCK(obj);
3264 }
3265 
3266 /*
3267  * vfs_page_set_valid:
3268  *
3269  *	Set the valid bits in a page based on the supplied offset.   The
3270  *	range is restricted to the buffer's size.
3271  *
3272  *	This routine is typically called after a read completes.
3273  */
3274 static void
3275 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3276 {
3277 	vm_ooffset_t soff, eoff;
3278 
3279 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3280 	/*
3281 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3282 	 * page boundry or cross the end of the buffer.  The end of the
3283 	 * buffer, in this case, is our file EOF, not the allocation size
3284 	 * of the buffer.
3285 	 */
3286 	soff = off;
3287 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3288 	if (eoff > bp->b_offset + bp->b_bcount)
3289 		eoff = bp->b_offset + bp->b_bcount;
3290 
3291 	/*
3292 	 * Set valid range.  This is typically the entire buffer and thus the
3293 	 * entire page.
3294 	 */
3295 	if (eoff > soff) {
3296 		vm_page_set_validclean(
3297 		    m,
3298 		   (vm_offset_t) (soff & PAGE_MASK),
3299 		   (vm_offset_t) (eoff - soff)
3300 		);
3301 	}
3302 }
3303 
3304 /*
3305  * This routine is called before a device strategy routine.
3306  * It is used to tell the VM system that paging I/O is in
3307  * progress, and treat the pages associated with the buffer
3308  * almost as being PG_BUSY.  Also the object paging_in_progress
3309  * flag is handled to make sure that the object doesn't become
3310  * inconsistant.
3311  *
3312  * Since I/O has not been initiated yet, certain buffer flags
3313  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3314  * and should be ignored.
3315  */
3316 void
3317 vfs_busy_pages(struct buf *bp, int clear_modify)
3318 {
3319 	int i, bogus;
3320 	vm_object_t obj;
3321 	vm_ooffset_t foff;
3322 	vm_page_t m;
3323 
3324 	if (!(bp->b_flags & B_VMIO))
3325 		return;
3326 
3327 	obj = bp->b_bufobj->bo_object;
3328 	foff = bp->b_offset;
3329 	KASSERT(bp->b_offset != NOOFFSET,
3330 	    ("vfs_busy_pages: no buffer offset"));
3331 	vfs_setdirty(bp);
3332 	VM_OBJECT_LOCK(obj);
3333 retry:
3334 	vm_page_lock_queues();
3335 	for (i = 0; i < bp->b_npages; i++) {
3336 		m = bp->b_pages[i];
3337 
3338 		if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3339 			goto retry;
3340 	}
3341 	bogus = 0;
3342 	for (i = 0; i < bp->b_npages; i++) {
3343 		m = bp->b_pages[i];
3344 
3345 		if ((bp->b_flags & B_CLUSTER) == 0) {
3346 			vm_object_pip_add(obj, 1);
3347 			vm_page_io_start(m);
3348 		}
3349 		/*
3350 		 * When readying a buffer for a read ( i.e
3351 		 * clear_modify == 0 ), it is important to do
3352 		 * bogus_page replacement for valid pages in
3353 		 * partially instantiated buffers.  Partially
3354 		 * instantiated buffers can, in turn, occur when
3355 		 * reconstituting a buffer from its VM backing store
3356 		 * base.  We only have to do this if B_CACHE is
3357 		 * clear ( which causes the I/O to occur in the
3358 		 * first place ).  The replacement prevents the read
3359 		 * I/O from overwriting potentially dirty VM-backed
3360 		 * pages.  XXX bogus page replacement is, uh, bogus.
3361 		 * It may not work properly with small-block devices.
3362 		 * We need to find a better way.
3363 		 */
3364 		pmap_remove_all(m);
3365 		if (clear_modify)
3366 			vfs_page_set_valid(bp, foff, i, m);
3367 		else if (m->valid == VM_PAGE_BITS_ALL &&
3368 		    (bp->b_flags & B_CACHE) == 0) {
3369 			bp->b_pages[i] = bogus_page;
3370 			bogus++;
3371 		}
3372 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3373 	}
3374 	vm_page_unlock_queues();
3375 	VM_OBJECT_UNLOCK(obj);
3376 	if (bogus)
3377 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3378 		    bp->b_pages, bp->b_npages);
3379 }
3380 
3381 /*
3382  * Tell the VM system that the pages associated with this buffer
3383  * are clean.  This is used for delayed writes where the data is
3384  * going to go to disk eventually without additional VM intevention.
3385  *
3386  * Note that while we only really need to clean through to b_bcount, we
3387  * just go ahead and clean through to b_bufsize.
3388  */
3389 static void
3390 vfs_clean_pages(struct buf *bp)
3391 {
3392 	int i;
3393 	vm_ooffset_t foff, noff, eoff;
3394 	vm_page_t m;
3395 
3396 	if (!(bp->b_flags & B_VMIO))
3397 		return;
3398 
3399 	foff = bp->b_offset;
3400 	KASSERT(bp->b_offset != NOOFFSET,
3401 	    ("vfs_clean_pages: no buffer offset"));
3402 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3403 	vm_page_lock_queues();
3404 	for (i = 0; i < bp->b_npages; i++) {
3405 		m = bp->b_pages[i];
3406 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3407 		eoff = noff;
3408 
3409 		if (eoff > bp->b_offset + bp->b_bufsize)
3410 			eoff = bp->b_offset + bp->b_bufsize;
3411 		vfs_page_set_valid(bp, foff, i, m);
3412 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3413 		foff = noff;
3414 	}
3415 	vm_page_unlock_queues();
3416 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3417 }
3418 
3419 /*
3420  *	vfs_bio_set_validclean:
3421  *
3422  *	Set the range within the buffer to valid and clean.  The range is
3423  *	relative to the beginning of the buffer, b_offset.  Note that b_offset
3424  *	itself may be offset from the beginning of the first page.
3425  *
3426  */
3427 
3428 void
3429 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3430 {
3431 	int i, n;
3432 	vm_page_t m;
3433 
3434 	if (!(bp->b_flags & B_VMIO))
3435 		return;
3436 	/*
3437 	 * Fixup base to be relative to beginning of first page.
3438 	 * Set initial n to be the maximum number of bytes in the
3439 	 * first page that can be validated.
3440 	 */
3441 
3442 	base += (bp->b_offset & PAGE_MASK);
3443 	n = PAGE_SIZE - (base & PAGE_MASK);
3444 
3445 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3446 	vm_page_lock_queues();
3447 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3448 		m = bp->b_pages[i];
3449 		if (n > size)
3450 			n = size;
3451 		vm_page_set_validclean(m, base & PAGE_MASK, n);
3452 		base += n;
3453 		size -= n;
3454 		n = PAGE_SIZE;
3455 	}
3456 	vm_page_unlock_queues();
3457 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3458 }
3459 
3460 /*
3461  *	vfs_bio_clrbuf:
3462  *
3463  *	clear a buffer.  This routine essentially fakes an I/O, so we need
3464  *	to clear BIO_ERROR and B_INVAL.
3465  *
3466  *	Note that while we only theoretically need to clear through b_bcount,
3467  *	we go ahead and clear through b_bufsize.
3468  */
3469 
3470 void
3471 vfs_bio_clrbuf(struct buf *bp)
3472 {
3473 	int i, j, mask = 0;
3474 	caddr_t sa, ea;
3475 
3476 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3477 		clrbuf(bp);
3478 		return;
3479 	}
3480 
3481 	bp->b_flags &= ~B_INVAL;
3482 	bp->b_ioflags &= ~BIO_ERROR;
3483 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3484 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3485 	    (bp->b_offset & PAGE_MASK) == 0) {
3486 		if (bp->b_pages[0] == bogus_page)
3487 			goto unlock;
3488 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3489 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3490 		if ((bp->b_pages[0]->valid & mask) == mask)
3491 			goto unlock;
3492 		if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3493 		    ((bp->b_pages[0]->valid & mask) == 0)) {
3494 			bzero(bp->b_data, bp->b_bufsize);
3495 			bp->b_pages[0]->valid |= mask;
3496 			goto unlock;
3497 		}
3498 	}
3499 	ea = sa = bp->b_data;
3500 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3501 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3502 		ea = (caddr_t)(vm_offset_t)ulmin(
3503 		    (u_long)(vm_offset_t)ea,
3504 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3505 		if (bp->b_pages[i] == bogus_page)
3506 			continue;
3507 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3508 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3509 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3510 		if ((bp->b_pages[i]->valid & mask) == mask)
3511 			continue;
3512 		if ((bp->b_pages[i]->valid & mask) == 0) {
3513 			if ((bp->b_pages[i]->flags & PG_ZERO) == 0)
3514 				bzero(sa, ea - sa);
3515 		} else {
3516 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3517 				if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3518 				    (bp->b_pages[i]->valid & (1 << j)) == 0)
3519 					bzero(sa, DEV_BSIZE);
3520 			}
3521 		}
3522 		bp->b_pages[i]->valid |= mask;
3523 	}
3524 unlock:
3525 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3526 	bp->b_resid = 0;
3527 }
3528 
3529 /*
3530  * vm_hold_load_pages and vm_hold_free_pages get pages into
3531  * a buffers address space.  The pages are anonymous and are
3532  * not associated with a file object.
3533  */
3534 static void
3535 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3536 {
3537 	vm_offset_t pg;
3538 	vm_page_t p;
3539 	int index;
3540 
3541 	to = round_page(to);
3542 	from = round_page(from);
3543 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3544 
3545 	VM_OBJECT_LOCK(kernel_object);
3546 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3547 tryagain:
3548 		/*
3549 		 * note: must allocate system pages since blocking here
3550 		 * could intefere with paging I/O, no matter which
3551 		 * process we are.
3552 		 */
3553 		p = vm_page_alloc(kernel_object,
3554 			((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3555 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3556 		if (!p) {
3557 			atomic_add_int(&vm_pageout_deficit,
3558 			    (to - pg) >> PAGE_SHIFT);
3559 			VM_OBJECT_UNLOCK(kernel_object);
3560 			VM_WAIT;
3561 			VM_OBJECT_LOCK(kernel_object);
3562 			goto tryagain;
3563 		}
3564 		p->valid = VM_PAGE_BITS_ALL;
3565 		pmap_qenter(pg, &p, 1);
3566 		bp->b_pages[index] = p;
3567 	}
3568 	VM_OBJECT_UNLOCK(kernel_object);
3569 	bp->b_npages = index;
3570 }
3571 
3572 /* Return pages associated with this buf to the vm system */
3573 static void
3574 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3575 {
3576 	vm_offset_t pg;
3577 	vm_page_t p;
3578 	int index, newnpages;
3579 
3580 	from = round_page(from);
3581 	to = round_page(to);
3582 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3583 
3584 	VM_OBJECT_LOCK(kernel_object);
3585 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3586 		p = bp->b_pages[index];
3587 		if (p && (index < bp->b_npages)) {
3588 			if (p->busy) {
3589 				printf(
3590 			    "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3591 				    (intmax_t)bp->b_blkno,
3592 				    (intmax_t)bp->b_lblkno);
3593 			}
3594 			bp->b_pages[index] = NULL;
3595 			pmap_qremove(pg, 1);
3596 			vm_page_lock_queues();
3597 			vm_page_unwire(p, 0);
3598 			vm_page_free(p);
3599 			vm_page_unlock_queues();
3600 		}
3601 	}
3602 	VM_OBJECT_UNLOCK(kernel_object);
3603 	bp->b_npages = newnpages;
3604 }
3605 
3606 /*
3607  * Map an IO request into kernel virtual address space.
3608  *
3609  * All requests are (re)mapped into kernel VA space.
3610  * Notice that we use b_bufsize for the size of the buffer
3611  * to be mapped.  b_bcount might be modified by the driver.
3612  *
3613  * Note that even if the caller determines that the address space should
3614  * be valid, a race or a smaller-file mapped into a larger space may
3615  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3616  * check the return value.
3617  */
3618 int
3619 vmapbuf(struct buf *bp)
3620 {
3621 	caddr_t addr, kva;
3622 	vm_prot_t prot;
3623 	int pidx, i;
3624 	struct vm_page *m;
3625 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3626 
3627 	if (bp->b_bufsize < 0)
3628 		return (-1);
3629 	prot = VM_PROT_READ;
3630 	if (bp->b_iocmd == BIO_READ)
3631 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3632 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3633 	     addr < bp->b_data + bp->b_bufsize;
3634 	     addr += PAGE_SIZE, pidx++) {
3635 		/*
3636 		 * Do the vm_fault if needed; do the copy-on-write thing
3637 		 * when reading stuff off device into memory.
3638 		 *
3639 		 * NOTE! Must use pmap_extract() because addr may be in
3640 		 * the userland address space, and kextract is only guarenteed
3641 		 * to work for the kernland address space (see: sparc64 port).
3642 		 */
3643 retry:
3644 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3645 		    prot) < 0) {
3646 			vm_page_lock_queues();
3647 			for (i = 0; i < pidx; ++i) {
3648 				vm_page_unhold(bp->b_pages[i]);
3649 				bp->b_pages[i] = NULL;
3650 			}
3651 			vm_page_unlock_queues();
3652 			return(-1);
3653 		}
3654 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3655 		if (m == NULL)
3656 			goto retry;
3657 		bp->b_pages[pidx] = m;
3658 	}
3659 	if (pidx > btoc(MAXPHYS))
3660 		panic("vmapbuf: mapped more than MAXPHYS");
3661 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3662 
3663 	kva = bp->b_saveaddr;
3664 	bp->b_npages = pidx;
3665 	bp->b_saveaddr = bp->b_data;
3666 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3667 	return(0);
3668 }
3669 
3670 /*
3671  * Free the io map PTEs associated with this IO operation.
3672  * We also invalidate the TLB entries and restore the original b_addr.
3673  */
3674 void
3675 vunmapbuf(struct buf *bp)
3676 {
3677 	int pidx;
3678 	int npages;
3679 
3680 	npages = bp->b_npages;
3681 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3682 	vm_page_lock_queues();
3683 	for (pidx = 0; pidx < npages; pidx++)
3684 		vm_page_unhold(bp->b_pages[pidx]);
3685 	vm_page_unlock_queues();
3686 
3687 	bp->b_data = bp->b_saveaddr;
3688 }
3689 
3690 void
3691 bdone(struct buf *bp)
3692 {
3693 
3694 	mtx_lock(&bdonelock);
3695 	bp->b_flags |= B_DONE;
3696 	wakeup(bp);
3697 	mtx_unlock(&bdonelock);
3698 }
3699 
3700 void
3701 bwait(struct buf *bp, u_char pri, const char *wchan)
3702 {
3703 
3704 	mtx_lock(&bdonelock);
3705 	while ((bp->b_flags & B_DONE) == 0)
3706 		msleep(bp, &bdonelock, pri, wchan, 0);
3707 	mtx_unlock(&bdonelock);
3708 }
3709 
3710 int
3711 bufsync(struct bufobj *bo, int waitfor, struct thread *td)
3712 {
3713 
3714 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, td));
3715 }
3716 
3717 void
3718 bufstrategy(struct bufobj *bo, struct buf *bp)
3719 {
3720 	int i = 0;
3721 	struct vnode *vp;
3722 
3723 	vp = bp->b_vp;
3724 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3725 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3726 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3727 	i = VOP_STRATEGY(vp, bp);
3728 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3729 }
3730 
3731 void
3732 bufobj_wref(struct bufobj *bo)
3733 {
3734 
3735 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3736 	BO_LOCK(bo);
3737 	bo->bo_numoutput++;
3738 	BO_UNLOCK(bo);
3739 }
3740 
3741 void
3742 bufobj_wdrop(struct bufobj *bo)
3743 {
3744 
3745 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3746 	BO_LOCK(bo);
3747 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3748 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3749 		bo->bo_flag &= ~BO_WWAIT;
3750 		wakeup(&bo->bo_numoutput);
3751 	}
3752 	BO_UNLOCK(bo);
3753 }
3754 
3755 int
3756 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3757 {
3758 	int error;
3759 
3760 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3761 	ASSERT_BO_LOCKED(bo);
3762 	error = 0;
3763 	while (bo->bo_numoutput) {
3764 		bo->bo_flag |= BO_WWAIT;
3765 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3766 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3767 		if (error)
3768 			break;
3769 	}
3770 	return (error);
3771 }
3772 
3773 #include "opt_ddb.h"
3774 #ifdef DDB
3775 #include <ddb/ddb.h>
3776 
3777 /* DDB command to show buffer data */
3778 DB_SHOW_COMMAND(buffer, db_show_buffer)
3779 {
3780 	/* get args */
3781 	struct buf *bp = (struct buf *)addr;
3782 
3783 	if (!have_addr) {
3784 		db_printf("usage: show buffer <addr>\n");
3785 		return;
3786 	}
3787 
3788 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3789 	db_printf(
3790 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3791 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n",
3792 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3793 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno);
3794 	if (bp->b_npages) {
3795 		int i;
3796 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3797 		for (i = 0; i < bp->b_npages; i++) {
3798 			vm_page_t m;
3799 			m = bp->b_pages[i];
3800 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3801 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3802 			if ((i + 1) < bp->b_npages)
3803 				db_printf(",");
3804 		}
3805 		db_printf("\n");
3806 	}
3807 }
3808 #endif /* DDB */
3809